# David Monson: every case study, in full > I've built for 15+ companies, from startups to Symantec. I set the AI strategy, ship the product myself, and train your people until every one of them uses it. Full time or fractional. Source: https://davidmonson.ai. The same words as the pages, without the layout. --- # Lightspeed https://davidmonson.ai/work/lightspeed ## Every designer, PM and engineer moved onto the AI platform I built. In 2026 I took SchoolAI's product org off Figma and into real code. Lightspeed reached 100% adoption, the team built 36 prototypes in its first month, and the company dropped a $45K Figma contract. Here's how I built it, and the decisions that made it stick. Lightspeed, at SchoolAI. Head of Design. I set the direction and built it with Claude Code. 2026. - 100%: Of designers, PMs and engineers on it - 36: Prototypes in its first month - $45K: Figma contract the company dropped - 57 to 1: Railway projects. One per prototype became one platform ### What happened. Every designer, PM and engineer moved onto it. The team built 36 prototypes in the first month. The company dropped Figma and its $45K contract. It's still how SchoolAI builds product. These numbers come from the system itself, as of early October 2026. ### Designers drew pictures while the product got built in code. Our designers worked in Figma and our engineers worked in code, so every handoff lost detail and time. We'd started prototyping with Claude Code, and it was fast. It also made a new mess. Each prototype ran in its own container, on its own Railway project. We got to about 57 of them. A running container costs money, so a prototype was set to expire after 24 hours, and any feedback went with it. Research took one path and comments took another. Nothing made anyone write down what would count as working. In June 2026 I put two lines on the design roadmap. Share Claude prototypes. And have designers build at production level in Claude Code, so we hand engineers finished UI instead of a design artifact. Lightspeed is how those two lines got done. A static build is a folder of files. Storing it costs pennies. So a prototype never has to die. That one change in the math made the rest possible: permanent links, version history, comments pinned to the exact build someone was looking at, recorded research you could come back to. None of that works while a prototype disappears overnight. ### Three repos, one system. It's production software for internal users, and it holds the line that keeps unreleased design work away from competitors. So I built it to that standard. ### An agent in your terminal can read all of it. All the feedback was already in Postgres and on a screen. But the agent in a designer's terminal couldn't reach it, so the designer had to paste it in. I wanted anyone to be able to ask Claude "what's the latest feedback?" and get crit, comments and tests back from any prototype. I chose MCP over a command-line tool for one reason. Claude Code, Codex and Cursor all speak it, and all three read a config file committed to the repo. Register the server once, and every session anyone starts can already see the feedback. Then I added the part I care about most. Lots of feedback is poor, and I didn't want agents building on all of it. So the tool that writes a fix brief refuses to run until a person has recorded their choices. ### Then I made the agents argue with us. Most of what goes wrong with a prototype goes wrong before any code exists. Nobody wrote a brief. Nobody argued about the problem. Six weeks later there's a demo, everyone nods, and no one can say if it worked because nothing was ever claimed. So I wrote a set of adversarial skills that push hard on the problem, the solution, the customer and the business value. They come before the build, and they refuse to write code while they run. Every push borrows a real test from a well-known operator or investor, and says whose test it is, so a PM takes the second push seriously. ### Every build gets checked against the design system. Agents write fast and drift fast. So every build runs the design system's analyzer against what changed and attaches a report: a chip beside the brief, a count on each variant, a page of every violation, and the same data over MCP. One button copies a fix prompt for Claude. The self-healing part happens during the build. A separate agent with no build context reviews the diff in three passes: is this the right model for the person, does the screen hold up to a crit, does it match the system. Every finding goes back and gets fixed, then a fresh agent audits again. It stops at three passes and tells the owner what's left. That isolation matters, because the agent that took a shortcut has already talked itself into it. I also record every scan, so we can see themes across prototypes. A mistake in nine prototypes is a documentation problem. The same count in one is a branch. Read these numbers as a starting baseline from one month. ### Then I went after everything we still needed Figma for. A team won't leave a tool because you tell them to. They leave when the new one does the job. Designers needed to inspect and measure. Engineers needed the spec. Everyone needed to look at the same thing together. Once all of that lived in Lightspeed, there was no reason to keep paying for it, and the company dropped Figma and its $45K contract. ### My first crit tool failed in front of the team. We do design crit often, and I wanted it inside the platform. The first version copied the presenter's clicks onto everyone else's browser. It demoed well. In a real crit, people landed on different screens, menus didn't open, and clicks hit the wrong thing. The design itself was wrong. A prototype has its own timers, animations and layout, and everyone's window is a different size, so one missed click puts someone off for good. So I threw out the model. The presenter's browser is now the only one that changes anything. Everyone else watches a live stream of it, with a comment layer on top. We ran a real crit on the new version on September 17, and it failed in ten places at once. We reproduced each one in a two-browser test harness and fixed it. The model held. ### How a whole org actually moved. Adoption is the job. Nobody cares about your platform until it's the easy way to do their work, so I kept taking steps away. ### What people built on it. 36 prototypes in the first month, and most weren't mine. At the start of October, five of them were. That's what I wanted. The platform was my work. The prototypes belonged to the team. All of these are explorations with mock data, so don't read them as shipped features. ### It moved quickly. ### How I'd run this at another company. You don't need my code. You need the same few moves, in the right order. I'd build it with your people, in your repos, and they'd own it when I'm done. --- # The Referee https://davidmonson.ai/work/referee ## Weak performers got through 11 times in 12. A separate AI referee cut that to 1 in 56. At Voltage, people practice hard conversations out loud with AI voice characters. When one model both played the character and decided who passed, it let people through who hadn't earned it. So I split the job between a character, a separate referee and plain code. I built it with AI agents, 721 commits in 60 days, and it's rolling out soon to a 130-person customer. Voltage, a voice-AI startup. Founder. In testing, rolling out soon to a 130-person customer. - 11 of 12: Times a weak performer got through when one model did both jobs - 1 of 56: Same kind of performer, across all seven steps, with a separate referee - 93.9%: Referee verdicts that matched the labeled answers - 0: False passes in the final eval ### The actor can't be the judge. Splitting the jobs fixed it. With one model doing both jobs, a nervous, unprepared simulated person got through 11 times in 12. With a separate referee, the same kind of performer got through 1 time in 56, across all seven steps. In the final eval, the referee matched the labeled answers 93.9% of the time, with 0 false passes. I can't show you its prompts or test data, so this page is the method and the measurements. Every pass rate here comes from my test lab, with simulated people. None of it is customer data. ### A character who wants to please you can't grade you. People practice hard conversations out loud with AI voice characters. The character has to push back the way a real person would, and only give way if you've earned it. At first, one model did both jobs. It played the character, and its prompt also told it when you had done enough to pass. On October 19, three separate investigations came back with the same answer. The character recited the rule back to the person. It acted out the end of the success path by itself. And it let weak performers through. ### The usual fixes made it worse, so the model wasn't the lever. I tried the normal fix first. I rewrote the rule as a clear checklist. It got worse. A nervous, unprepared simulated person went from getting through 7 times in 12 to 11 times in 12. A checklist inside the prompt just gave the character more lines to say out loud. Then I raced other models on the same test. Every one let that person through, and the best instruction follower did it most. That told me the model wasn't the lever. The job was wrong. ### So I split it into three parts, and only two of them are AI. One model plays the character. A second model reads each line and judges it. Plain code keeps score and decides what happens next. The character is told only where things stand, in a short note, and it never sees the rules. A pass and a score are now separate things. In my words from that night: "You should be able to complete a thing even if you don't check every single thing off. But the score and the feedback afterwards should absolutely be rigorous on what was missed." ### An eval chose the model, and my first prompt failed it. I built the test before I picked a winner. It has 176 lines, from simulated conversations, one real conversation of my own, and a few lines written to be hard: a long intro, rude talk, and an instruction hidden in the conversation. Each line was labeled against the contract. Each model then ran every conversation on its own ledger, the way the live system does. The first prompt scored 64% on Haiku 5.5 and 59% on Haiku 4.5. They let 5 and 13 people through who should have been stopped. The cause was a single word. "Partial" meant "what this line added" to the labels and "where it stands now" to the models. One question also had five parts that I had never spelled out. So I fixed the prompt before shopping for a model. Partial now means where it stands. The parts are numbered. It carries forward what's still missing. And it has to decide whether the question was even asked before it can call something a dodge. Only then did the race mean anything. Replaying those same conversations on the labeled verdicts gave me the shape I wanted: by-the-book got through 6 of 7, average 1 of 4, and weak 0 of 13. That includes a call of my own that had wrongly passed. ### The cheap model never let anyone through wrongly. I still passed on it. Haiku 5.5 was faster and cost about a nineteenth as much. It had zero false passes. But it stopped 2 to 4 conversations that should have gone through, and a person doing it right doesn't deserve that friction. Sonnet 5.5 misses my 1.5 second guide by about 0.3 seconds, and that's fine, because of the one-reply lag. Later I loosened the standard so that the gist of an answer counts, in the person's own words, and the eval re-ran with new labels: 95.7% of verdicts agreed, and 33 of 36 outcomes matched. Adding off-track judging later put agreement at 89.8%, against 91.4% for the same model just before the change. That's one or two lines, and I wrote it down. It caught 13 of 13 off-track lines, with 0 of 186 honest lines flagged. Every model in the product is a swappable line item. Evals decide when one gets swapped. ### The rule: quality holds, and either cost or latency drops. A person approves every swap. I keep a switch log with the receipts. A weekly job that runs this automatically is designed, and I'm not claiming it's running. The character's model used to be one constant in the code. Now it's a setting, resolved per conversation, and recorded on each one. ### Each step is a contract, and plain code keeps the score. This is the method. The prompts and test sets stay private, but the shape is the part any team can copy. ### It runs on vendors I can swap, with the referee on Claude. ### Prompts are production code, so they get the same guardrails. "We lose quality when it's too long" was my own note in September. So the size limit is a test. A prompt over its ceiling fails the build, offline, with no API key. The ceilings only go down. Judge changes run against a recorded baseline. A score that moves more than the tolerance, crosses a band, or gains or loses a flag fails the check. The rule on that file: regenerate it only after you understand why the score moved, never to make the check pass. ### A note from a tester can become a tested fix. Testers can talk or type feedback from anywhere in the app. Two rules hold the pipeline up. A submission is never lost because an AI call failed, so the first step just stores what was said, and everything after it is a queue. And the loop closes: the person who reported the problem hears that it changed because they said so. ### I measured how easy it was before I made it harder. I built a lab of simulated people who play set skill levels against the live character. Before the referee, 168 simulated conversations cost about $6. By-the-book performers won 80% of the time, average ones 77%, and weak ones 52%. The verdict in my notes: far too easy. The character also led the person in 15% of its turns and scored 3.0 of 5 for realism. Then I wrote down what hard means. A hard conversation is one you can win if you earn it: about 7 in 10 for by-the-book, 3 in 10 for average, 1 in 10 or fewer for weak. That's friction you can overcome, with no dice roll, and no cold character who can't be won over. ### Then I told it to stop chasing the number. Average performers were passing 12 of 12 at two steps, against a bar of 30 to 60%. The easy move was to make the character harder until the number looked right. I said no. At those points in the conversation the character has already agreed to the premise, so it should be easy to finish, and the judge is what grades quality. The same week, I dropped the idea of several character types. In a side by side, the one lenient type let 3 of 3 average performers through against 1 of 9 for the others. I wanted one consistent bar with variety in the wording, so that's what it became. ### Cost is a product requirement, so I measure it per loop. Voice is about 90% of the cost. So I wrote down the exit before I needed it: if all-in voice cost passes a set ceiling, I pull forward a self-hosted path on open models. The kill threshold is in the architecture rules, and I don't have to win an argument about it later. ### A second pipeline: 28 videos, no film crew. I regenerated a set of 28 existing videos as avatar video with HeyGen, with an agent driving the browser. It made 69.8 finished minutes, at about $2.40 a finished minute. Cash paid that cycle was under $100, with about 7 hours of wall clock over two days, mostly waiting on renders, and about 2 minutes of my time per video. ### Plenty failed first. Here it is, in the order it hurt. ### What I'd tell any team building AI that has to be trusted. ### It's rolling out soon to a 130-person customer. Here's what I can't claim yet. It's in testing, and it's rolling out soon to a 130-person customer. The referee is on across all seven steps in that test build. Here's what I can't claim yet. I have no usage numbers, no outcome numbers, and no proof yet that a good score here predicts anything in real life. The lab uses simulated people. The next real test is comparing the judge against human grades on about 30 conversations. If those disagree, the judge has more work to do, and I'd rather find that out from a measurement. --- # Mastery Arcade https://davidmonson.ai/work/mastery ## One person and a team of AI agents built a live arcade of learning games in five months. Mastery Arcade is two games on one learning engine: Times Machine, with 17 levels and 4 boss fights for your times tables, and the Lab, with 22 hands-on 3D science experiments. I built it alone from May 30, 1,069 commits, for web, iPhone, iPad and Android from one codebase. Mastery Arcade. Founder and only builder. Product, design, code, AI media, growth. May 2026 to now. Live on the web, App Store launch planned for Nov 3, 2026. - 1,069: Commits on main, May 30 to Oct 31, all under my name - 1: Person building it, directing AI agents - 22: Hands-on 3D experiments in the Lab - 3: Platforms from one codebase: web, iOS, Android ### Two games, three platforms and 1,069 commits, from one builder. Five months, May 30 to October 31. I decided what to build and made the taste calls. AI agents wrote most of the code. The second game, the Lab, reached production with 11 experiments built in one day. These count what exists in the repo today. I'm leaving out reach and revenue until there's something to report. The web version is live, and the App Store launch is planned for November 3. ### Games you want to keep playing, with real practice inside. Kids and adults practice facts with flashcard apps that bore them, or with games that teach very little. Mastery Arcade is my answer: a growing set of learning games, built on one learning engine, with the feel of a real arcade. The first game is Times Machine. A giant machine has come to optimize Earth, and the only way to beat it is multiplication. You practice each table, earn a medal, then fight bosses named BYTE, MEGA and GIGA before facing the Times Machine itself. The second is the Lab, where you pick up a question like "how do planes fly?" and test it with your hands in a 3D room. One family pass covers everything, with up to six Players. It started on May 30 as one multiplication app. By June 27 I had rebuilt it as a monorepo around a shared engine, so a second game wouldn't mean starting over. ### One engine, so the game can't drift away from the learning. Practice, level drills, every boss, the challenges, the first-run drone mission and the free speed check on the website all ask questions through the same state machine. Each surface sets its own knobs, like the retry on a miss or the clock. Nobody gets to write a second one, because a boss fight that grades differently from practice would quietly break the promise of the product. ### The product is the output. The way I run the agents is what I'd bring to your company. I don't write most of the code by hand. I decide what to build, I make the taste calls, and I check the result by using it on a real phone. Agents do the rest. That only works with a system around them, and most of that system came from something breaking. On September 2, I filed nine epics and about 108 issues and let the agents go. They started around 6:15 pm and were still merging past 3 am. Over that four-day arc, 47 PRs landed, along with one production incident and one forced hosting upgrade. ### A character pipeline, down to about 25 minutes. I make and approve the character art myself in Midjourney. Agents run everything after that: Meshy turns the image into a 3D model, it gets rigged and animated, and it lands in the game as a GLB file in Three.js. I proved the whole path on one pilot, VEX, on August 23: 62 credits and about 25 minutes. The steps are simple to list. The rules around them are what cost me. They're written into the pipeline doc, each with the measurement that earned it. ### I directed the films like films, then measured what the models really do. Each boss has its own encounter: it arrives, taunts you, fights, and falls apart when you win. I made stills in Higgsfield with GPT Image 2, then video with Seedance. A captain named Reyes talks to the player between fights. "Identity is won or lost in the still" became a rule, because the video reproduces every flaw in its first frame. Every prompt and start frame for the four encounters took about 2,700 credits of directed work. Then I found out it all lived in a temporary folder the operating system could wipe. So now a ledger records each render's exact inputs. A vertical version of a film is a loop over that ledger, not a dig through old chats. ### The biggest sound fix was a default nobody had chosen. The soundtrack is my own Suno tracks, switched by scene. Reyes' voice is one I designed in ElevenLabs, cloned into Higgsfield, with a vibe note on every line. One lesson was bigger than audio. Her voice ran on one engine for a month until I asked why we weren't using ElevenLabs through Higgsfield. Side by side, the same intro ran 24.85 seconds against 18.24, and a laugh got performed where before it was spoken as syllables. Nobody had chosen the old engine. It was a default no one had decided, and that went into the brief. ### Then I built a second product in about a day. In mid-October I pointed the same system at physics. The first 11 experiments reached production on October 15, built in one day by about 15 parallel agents working from shared briefs for UX and art. By the end of October there were 22: smoke tunnels, Build an Atom, a roller coaster, orbits you can crash a moon through. Each one has a question, things to try, clues and a short quiz, and you earn a patch. Your own pilot from Times Machine shows up in eight of them. These clips are real captures of the app. ### A real app on Apple and Google, from one codebase. I was an iOS developer before I was a designer, so I knew the danger: an app that's only a website gets rejected. The apps are Capacitor shells around the live product, with native features a reviewer can find and use. Most of the store work was paperwork. I turned that into tested files, so the listing can't drift from the code. ### Money gets the strictest rules in the repo. I priced one family pass and one solo pass, sold on the web through Stripe. I planned Apple in-app purchase, then cancelled it. The iOS app is US-only for version one, and it opens Stripe in a browser sheet. The prices will change as new games arrive, so I built the system to make a stale price impossible to type. ### Growth is built into the product, ready for launch. All of this is in code. None of it has real volume behind it yet, so I can tell you what I built and nothing about how well it works. I also read my own weekly memo the same way: the first one found a real checkout that started and never finished, and showed a referral factor of zero. ### I turned the expensive lessons into skills that any agent can load. The biggest is a 153 KB skill on directing AI film, written from my own tests. A skill is a written playbook an agent reads before it starts a job. Some of these I wrote from scratch, and some I brought in and changed. Here's which is which. ### What it costs, as far as I can honestly say. My cost sheet has one rule: a blank means unknown, never zero. So it refuses to give a total while lines are missing. In September, I could price five of twelve lines. The rest stayed blank, so I don't have a true all-in number, and I won't estimate one here. ### Most of what broke was silent. Here's the list. Many of these were silent. Nothing errored, and nothing looked wrong until someone checked. ### Live on the web, with the App Store launch planned for November 3. The web version is live at masteryarcade.com. The App Store launch is planned for Tuesday, November 3. Apple hadn't received the app for review the last I checked, on October 31, so that date depends on review. If Apple runs late, the plan is to launch on the web on the 3rd and let the store follow. The launch is the real test, and I'll put the results here, good or bad. ### Everything it took, from Midjourney to the App Store API. --- # North Star https://davidmonson.ai/work/north-star ## One storyboard aligned Instawork on three years of work, and we got there faster than planned. In 29 days, the CEO, the VP Product and I turned a three-year product vision into 14 storyboard frames the whole company could retell, and teams started deciding for themselves. Over 18 months, pros grew from 400K to 4.5M and the app reached No. 1 in Business on the US App Store. Instawork. Head of Design, with the CEO and VP Product. 2021 to 2023. - 4.5M: Pros, up from 400K in 18 months - No. 1: Business chart on the US App Store - No. 3: All free apps, the same day - 29 days: Kickoff to all-hands ### What happened. Kickoff to all-hands took 29 days. After that, everyone was so aligned that teams just started deciding for themselves. The work was supposed to take three years, and we surprised ourselves with how fast it got done. The features the team designed and shipped drove the organic growth: 400K to 4.5M pros in 18 months, and No. 1 in Business on the US App Store. It's the crown jewel of my career, and a process I can run with anyone. ### Instawork was growing fast, and planning one quarter at a time. There was no shared picture of the product two years out. The CEO wanted to define what the product experience should be at the end of 2023. A written strategy alone wouldn't do it. People don't retell documents. They retell stories. ### I kept the core group to three, and we wrote before we drew. The core group was the CEO, the VP Product and me. I owned the user segmentation and the map of today's journey, with a rating of how well we served each stage. The CEO wrote a strawman. The VP Product looked at what others did uniquely well. Writing came first, because it forces complete sentences. Then pictures, because pictures travel. ### People retell stories, not documents. So we drew one, frame by frame. Inspired by Airbnb's Snow White exercise, we followed one pro and one partner through a five-star experience. The pro, Lisa, is a mom of two who needs more income and more flexibility. The partner, Rockefeller, owns a business and needs people he can trust. We wrote the narration. In 2021 a freelance illustrator drew 14 comic panels, and we printed them as posters. Here's the same story as I'd make it today, in 3D. Tap through it, or switch to the 2021 originals. ### Then we designed the screens the story promised. Once it was locked, we presented to the whole company and immediately started on the next generation of the app. I mapped out the information architecture and gave every team concrete direction. These prototypes made the story real enough to argue with. It was supposed to take three years. Everyone was so aligned that teams just started deciding for themselves, and we surprised ourselves with how fast it got done. ### The story kept showing up in the roadmap. Look at Lisa's story and you can read the product plan. A debit card that pays out instantly became the Instawork Card. Rewards for reliable pros became Top Pro. "Profile of record" kept appearing in company strategy documents long after the offsite. ### 400K to 4.5M pros in 18 months, and No. 1 in Business. The features the team designed and shipped drove the organic growth. Here's where it took us on the US App Store, May 10, 2022. ### Today I can direct the same story in an afternoon, in any style. In 2021, 14 panels meant hiring a freelance illustrator. Today I can direct a set myself in an afternoon, which changes when you'd reach for this exercise at all. The 3D set above is the one I'd use here. A different company might need something else, so here are the same three moments in other looks. Tap one. --- # SipsClub https://davidmonson.ai/work/sipsclub ## A new drink brand's store converted at 4% on day one, and 3.14% for the year. SipsClub needed a store, a story and a way to sell a product type nobody had a word for yet. I built all of it with no agency and no developer on staff, then ran the tech for a year as its fractional CTO: $149,299 in net sales, 1,522 orders, and more than 30% of sales from subscriptions. SipsClub. Fractional CTO. Strategy, design, code, copy, SEO, video, analytics. 2024 to 2025. - 4%: Conversion on launch day - 3.14%: Conversion across the whole first year - $149,299: Net sales in year one - 30%+: Of sales from subscriptions ### 4% on launch day, and above 3% for the whole first year. The store launched in June 2024 and converted at 4% on day one. Across the first year it held 3.14%, for $149,299 in net sales. Subscriptions grew past 30% of sales, and organic search brought in over $17,000. No agency team and no developer on staff. These numbers came from the systems on this page. ### A new kind of product, and no one to build the store. SipsClub makes drink mixes for families. At launch that meant a daily digestive drink nobody had a category for yet. A fruit-flavored protein drink and a caffeine drink followed, each in several flavors. The founders had the science and the recipes. They didn't have a website, a store, a tech stack or anyone who knew how to sell online. So I took it. Platform, design, code, copy, search, email, video, analytics. If it touched a customer on a screen, I built it. The store went live in June 2024. You can still see the version I launched on the Wayback Machine. It has since been redone. ### The first call: Shopify, not WooCommerce. This team needed something they could run without me in the room. I compared Shopify and WooCommerce on security, real cost, ease for non-technical people, and how reliable the apps were. I recommended Shopify, and the team went with it. Then I found a set of apps that could work together: flavor bundles, one-time and subscription orders, and a fulfillment partner with its own rules. That work decided whether the rest was possible at all. ### Mix-and-match subscriptions were the hardest part, and customers never saw it. This was the hardest part of the whole job. A customer could subscribe or buy once, and mix and match three flavors inside a single variety pack. Each of those choices had to reach the warehouse as the right items, every time, on every renewal. Shopify, Recharge, ShipStation and the fulfillment partner each had a different idea of what a product was. Nothing did this out of the box. I spent days with Recharge's own developers on our exact setup, then wrote custom Liquid to cover what was left. The customer saw none of it. They picked flavors and tapped a button. Subscriptions grew past 30% of sales. ### Every page was designed phone first, with Lighthouse scores of 90+ on mobile. I started from a Shopify theme chosen for flexibility and long-term support, then tuned every part of it to the brand. I designed for the phone first. ### I finished the packaging, and one band of color fixed a real problem. I recommended the design agency that ran the brand exercise, and I stayed close through it. That gave us the logo, colors and guidelines. When packaging still wasn't finished, I stepped in and finished it for every product, in Figma, including the label layouts the printers needed. My favorite fix is small. Customers couldn't tell flavors apart at a glance. So I added a color-coded flavor strip across every packet. A problem you could feel in the kitchen, solved with one band of color. ### You can't sell what people have no word for, so every page teaches first. You can't sell something people have no word for. So every product page was built to teach first, then ask. I wrote all of it, using ethos, pathos and logos on purpose: why trust us, why this matters to your family, what's actually in it. One detail did more than any headline. Some people look for a specific form of folate on a label, 5-MTHF, because of a gene variant called MTHFR. We used that form, and we said so. Most shoppers skipped right past it. The ones who knew recognized themselves and bought. It was a quiet "if you know, you know," and it converted. ### I directed the launch videos for about $500 each, where that work usually costs $6,000 to $8,000. I scripted and storyboarded the launch videos, chose locations, talent and props, and recruited a crew. The cost came to about $500 a video. Work of that quality usually costs $6,000 to $8,000. I planned where each one would live before we shot it: the homepage, product pages, Reels, TikTok, Facebook and email. They drew tens of thousands of views, and they kept selling long after launch. ### A launch reel is built for watch time, beat by beat. Here's the structure from one of my scripts, for the Support Sips launch on Instagram. The note at the top said to cut out every tiny pause, because watch time was the whole game. ### One TV segment, scripted to sell, brought in about $10,000. Most brands treat local TV as press. I treated it as a sales channel. For Good Things Utah, I scripted the segment topics, then designed the path from the couch to the checkout, with a QR code ready for people watching live. That one segment brought in about $10,000 in sales within a few days. I'd pushed for this kind of long-form appearance because it gave the founders' science background room to build trust. It became the best sales strategy the company had. ### We beat Panera for our own name. Panera's Unlimited Sip Club crowded every search for our name. A new brand could easily have disappeared behind it. So search was built into the site from the first day, not added later. We took No. 1 for "Sips Club" and won the Google Knowledge Panel. Organic search brought in over $17,000, about 11% of first-year sales. ### Add-to-cart tests lifted conversion 15 to 20%, and a cleaner checkout cut abandonment. Launching at 4% is good. Holding above 3% for a year took constant testing, with results read week by week. The add-to-cart tests lifted conversion 15 to 20%. A cleaner checkout layout cut cart abandonment by about 12%. ### I argued to keep the email pop-up. There was pressure to take it down. I kept it, and that decision became a sales channel. It captured 1,669 emails, which brought in $7,172.62 in tracked sales. All the email marketing that came after depends on that list. ### A year as fractional CTO, and no major bug reached the live store. For more than 12 months after launch, I was the company's fractional CTO. I fixed bugs, launched new products, tested after every app update, refreshed pages, read the numbers and changed course. No major bug ever reached the live store. New products became routine: clone the template, test every SKU path, refresh the pages, write the launch. That was the point of building it this way. --- # Boosted Safe https://davidmonson.ai/work/boosted ## Conversion up 83%, from 0.6% to 1.1%+, in two months. Boosted Safe sells a steel car safe disguised as a booster seat. From September 1 to November 1, I rebuilt everything a shopper touches and most of what they never see, using Claude Code and a crew of agents: 192 pieces of work, a home page cut from 60 MB to 14 MB, and email signups up 56%. Boosted Safe. Shopify Plus rebuild. Research, design, theme code, copy, SEO, AI search, testing, analytics, automation. September 1 to November 1, 2026. Numbers as read on November 1. - 0.6% to 1.1%+: Conversion, up about 83%. 1.1% is what I launched - +56%: Email signups, 48 to 75 a day. 3 days against the week before, not an A/B test - 3.0% to 14.1%: Orders with 2 or more safes: 15 of 501 over 4 weeks, then 12 of 85 in the deal's first 4 days. Small sample - 60 to 14 MB: Home page weight ### Conversion rose from 0.6% to 1.1%+ in two months. The store was at about 0.6% when I started, and I launched at 1.1%. Email signups rose 56%. Orders with two or more safes went from 3.0% to 14.1%, on a small sample, once a clear multi-safe deal replaced old joke discounts. The home page dropped from 60 MB to 14 MB. Conversion is the headline, and the next chapter explains how to read it while tests run. The rest comes from Shopify, Klaviyo, Search Console and Lighthouse, and each one says what it's measured against. ### 0.6% to 1.1%+, and why the live number reads lower while tests run. The store was at about 0.6% when I started. I launched at 1.1%. Long-term, the goal set at kickoff is 3.5%, and this is the first big step toward it. Since launch we've kept testing on live traffic. An A/B test holds part of the traffic on another version of the site, and when that version loses, it pulls the store-wide figure down for as long as it runs. We're about to end the tests, and then the store goes back to the winning version. ### The product was clever, and the old store buried it on a phone. Boosted Safe clips into the child-seat anchors already in your back seat, hides under a booster-seat cover, and opens with your finger. It's patented, and it's a great idea that's hard to explain in one glance on a phone. You can see the live store here. The store gets about 3,000 sessions a day, with about $3K a day going into Meta. 84% of sessions come from phones, and more than half arrive from Facebook and Instagram, mostly inside those apps' own browsers. The old site met them with a discount code in the top bar, a pop-up over everything, and a product page that asked you to read a diagram. I took on the whole storefront. My first rule to the agents, in my own words: "number one priority is we don't do anything to take down or affect the current site." Everything was built on a draft until the night it went live. ### Before designing anything, I read everything. Two weeks of this came before a single layout. It's why the site says what it says, in the order it says it. ### The hero stat was wrong, and the true one was better. The store said a break-in happened "every 36 seconds." When I went looking for the source, it was old and it was about stolen cars. FBI data gives one reported theft from a vehicle every 45 seconds. The Bureau of Justice Statistics finds police only hear about one in four. That works out to about one every 11 seconds. So the site now says one every 11 seconds, and prints the math and the citations right under it, on a dial. A claim about 60,000 stolen guns got the same treatment and became a sourced figure from Everytown. The true number turned out to be the stronger one. It usually pays to go and look. ### A checkout spike turned out to be phantom events. From August 26, the numbers looked like a breakthrough. On the same ad spend, add-to-carts doubled. It was tempting to credit Meta's optimization. Purchases, though, fell 12%. I traced the start to the day a theme was published. That publish was firing phantom cart and checkout events, and the pixel passed them to Ads Manager. It stopped on September 5. Without that trace, everyone would have spent months reading bad data as good news, and the baseline would have been wrong too. ### I built this with agents, and I stayed the one making the calls. Nearly all of this came out of one long Claude Code session, September 4 to November 1, with more than 500 messages from me. I gave direction, reviewed in real browsers, and approved anything that touched the live store. When I was away from the desk I drove it from my phone with Remote Control. Agents let one person cover what used to take a team. They also run ahead, get things wrong with confidence, and will happily make things up. So most of the craft here was in the setup: who plans, who writes, who checks, and what no one is allowed to do. ### Some things were simply off limits to the agents. Agents can do real work on a live store when some things are simply off limits. These were the limits. ### A claims gate caught a real error. Agents drafted the articles and the video pages. Nothing published until a claims check passed. That gate cut 69 spoken lines from transcripts and held 2 videos back. Then a writer agent, reading the transcript of the store's own pull test film, noticed that wording on a live page didn't match what the film shows. That's an error, and it's being corrected. The fix is written and committed, and it's on its way live. That's the case for agents with a gate: more eyes reading more of your own material than any person has time for, and a rule that says the record wins. Nothing invented ships. ### 19 looks and 14 rounds of feedback got a yes in about two days. I built homepage directions fast using one shared copy kit, so the owner was judging looks and not wording. Six became 19. They got consolidated into a 12-page prototype in 7 skins, with light and dark, all on a client review hub on Vercel with a short guide on how to review. The direction we landed on took its cue from onewheel.com: big type, real product, confident. Those 14 rounds of feedback happened in about two days, and the owner approved the look on September 20. ### A heavier logo, and a brand bible a phone can use. The old rounded wordmark looked soft for a steel safe. I drew a heavier squared mark and a Big Shoulders wordmark, full vector, and tested options side by side in the sizes they'd actually live in, down to a phone header. The brand bible is a phone-first site with copyable values, downloads and a gallery. Its colors and type read from the theme, so it can't drift away from the store. ### A custom store you can still upgrade. I planned to build on the Impact theme. Three research agents and a check of primary sources changed my mind, and I moved to Dawn 16, Shopify's own reference theme. Dawn stays untouched apart from three fenced hooks, and everything custom sits in a separate layer. When Shopify ships an upgrade, it can come in. On September 23 I asked how solid the build really was. That question led to written rules and a clean-up against them: inline styles went from 130 to 8, inline scripts from 3 to 0. ### Cold traffic won't read a paragraph, so the pages give it things to touch. Most of the traffic is cold, from a social ad, on a phone. A paragraph won't hold those people. So the home and landing pages are full of things to touch, each one making a point about the product. ### Six safes run from one section, so a change to one is a change to all. Every product page comes from one section and a seed script, so a change to one is a change to all. Six safes and a refurbished model, with the Elite pre-selected. The hard parts were pre-orders, financing and showing the free items without lying about any of them. The renders and studio photos came from the client's creative team. I chose them, wrote the words, composed the 16:9 tiles as HTML, and rendered them with Playwright. So a copy change re-renders a whole gallery. ### The pages nobody puts in a portfolio answer the question that stops a sale. A store is mostly these. They answer the question that stops a sale, and they're usually where a redesign runs out of patience. ### Every buyer gets a page that starts with their problem. A photographer and a parent want the same steel box for different reasons. So the landing pages open on the problem, then a calculator preset for that person. Ads could point at the right door. California needed its own story. The Golden Gate Elite is on the state's Department of Justice list of certified firearm safety devices, so I built a section that shows the listing and links you to check it yourself. ### Free items that work on pre-orders, and a deal that replaced a joke. Each safe comes with free items. Shopify's Cart Transform is the usual tool for that, and it fails on pre-orders, which is much of what this store sells. So I rejected it. Each free item is a real product, added automatically at $0 and matched to quantity, and a Flow closes the line at order. A warehouse note stops anyone picking extras. 16 cart checks cover it, and 202 of 209 site orders carried them. Then there were the discounts. I found 8 old "joke" discounts still running, with names like "The Ocho." They stacked, they hit products that should never be discounted, and they broke referral codes. I replaced them with a clear multi-safe deal: 25% off at 2 to 9 safes, 30% at 10 or more. It shipped by script with a dry run, a go, and an undo. The script won't retire the old discounts unless real carts have priced exactly right in the last two hours. ### "Shop starting at $32,079.53." A currency bug, fixed the same night. A screenshot came in through Messenger showing that on one of our pages. It was Jamaican dollars. Shopify Markets was converting prices for foreign visitors, and my big blue button was reading that out loud. Before changing anything, I checked 90 days of orders. All 1,394 were US orders in US dollars. So I fixed it in Markets that same night. A small thing, and the sort of thing that only turns up when someone is watching the store. ### I cut the home page from 60 MB to 14 MB. The old home page was a 60 MB download. On a slow phone, in a lab test, it took 36.3 seconds to load. Files moved to Shopify's CDN, images went to WebP with space reserved, and videos got poster stills and lazy loading. Each measure below is the median of 3 Lighthouse runs on October 12, old live theme against new draft, unless it says otherwise. The honest part: the lab performance score barely moved, 31 to 33. Blocking time comes from third-party apps, and those set the floor. I recommended cuts. Not all have happened. ### Publishing would have silently killed email capture. Five days before launch I had an agent compare the live theme and the draft, app by app. The live theme had 12 app embeds on. The draft had 2. App embeds switch on per theme, so 10 of them would have gone dark the moment we published, with no error. The only email capture on the site was one of them, so email signups would have gone to zero. Affiliate tracking and the dealer form would have died the same way. The audit also turned up a dead speed app still in the code and an upsell script registered five times. The audit gave every app a specific action to take before publishing. ### Day one was the weakest of 8 days. Then we fixed 14 things. The first 24 hours brought in $5,353 and 18 orders on 2,843 sessions, below all 7 days before it, which ran $5,803 to $13,724 and 20 to 32 orders. Add to cart was fine at 2.92%, the third best of the eight. The gap was at the end: 23% finished checkout against 26 to 42%. I wrote that up the same night. The likeliest cause was a shift in ad mix: ad visits to one Mini page tripled while the main Elite ad page got half its usual visits. And part of it was mine. A whole product had no link. I also wrote, "One day of 18 orders is mostly noise," and that the first real read was a week away. ### We lost our own test, and I said so. I tagged every cart, GA4 event and Clarity session with a site version and first-visit date, so any order could be split by test and by who it actually touched. Then we ran real A/B tests on live traffic through Shopify Rollouts, with about $3K a day of Meta spend behind it. Test 1 was a long-form "infomercial" home page that the whole team believed in. The plan came from scoring the current home page by appeal: 18 sections of logic, 13 of credibility, only 4 of emotion, with about 20 shop buttons and prices not appearing until section 13 of 37. It lost, 0.57% to 0.94%, 19 orders to 30, $7,242 to $11,336. I ended it early and wrote it up plainly. I'd also said too soon that it passed a 95% bar, and I withdrew that the next day. Shopify's own report said "not enough evidence." The useful bit: the long page got more shop taps even with the button hidden. So its parts moved onto the site one at a time, instead of all at once. ### The clear winner that turned out to be a tie. Test 2 hid the menu from first-time visitors. It read 1.6% against 0.87%, which looks great, on a day and a half of traffic. That's far too little, and I ended it to merge everything into one theme. Test 3 asked the same question with one setting as the only difference. It ran three times, because Rollouts froze a copy of the control and my updates only reached B. Final read: A at 1.02%, B at 0.78%, 49 orders to 39, $17,434 to $15,490. It looked like a win for A. An agent called it "the clear winner." But the menu only hides on a first visit, so anyone whose first visit came before the test saw the same site either way, and A had simply drawn more of those people by luck. A drew 18 orders from that group and B drew 11, a 7-order swing inside a 10-order gap. Take them out and it's 31 orders to 28, $11,133 to $10,917. B's buyers spent more, $390 an order against $359. My verdict moved from "close to proven" to 1 in 18, to 1 in 5, to about 1 in 4, a tie. I reported each correction as it happened. The home page keeps its menu. ### The pop-up that said "It ends today." It didn't. The old pop-up used false urgency. I replaced that line with "Lock it up instead." The new one is full screen with picture tiles for what you're protecting, including a handgun, a laptop, a wallet and a PSA 10 Charizard. The SMS step headline is a real review: "Wish I would have bought one sooner." It fits a phone without scrolling, with the button always in reach. Email signups rose 56%, from 48 to 75 a day, and SMS 2.5x, from 15 to 37 a day, comparing the 3 days after it went up with the week before. Klaviyo-attributed revenue went from $1.4K to $3.3K a day, which is attribution only. That's a short read and not an A/B test. Site conversion held steady, so it didn't hurt. Then I named the next leak. People who start now quit at the email step 61% of the time, up from 47%. And the SMS list grew by 89 against 112 phone-step completions. I can't explain that gap yet. ### Search is growing, and we are not #1 yet. That's the first line of the scorecard I wrote for the team. Category search is growing and nearly every page is indexed. Brand clicks and organic revenue fell. All of that is in the report, and all of it is here. It's early. These are 18-day windows, most of the SEO pages went live October 24 or later, and ad spend moved across the period. The big head terms sit at the bottom of page 1: "car safes" at 9.8, "car gun safe" at 10.9, and "console safe" not ranking at all. That's the starting line. ### People ask ChatGPT what to buy now. So I measured that too. I ran the same buyer questions through four AI engines. On questions that don't name a brand, Boosted Safe came up in 33 of 88 answers, and first in 17. Console Vault came up in 51 and Vaultek in 37. Gemini and Perplexity don't name the store at all on the 14 biggest terms, and Google AI Mode does on 12 of 14. Read these as a best case, since they ran in my own signed-in browser. An earlier run used a different engine, so there's no fair trend yet. The first true comparison is December 1. ### Black Friday, built so Christmas is a settings change. The sale runs all of November. I built it into the theme: name, percent, dates, color and countdown are all settings, with struck prices across every buy box and a preview by URL. Switching to Christmas means changing a few fields. I tested the discount rules with real carts and temporary products: the biggest discount wins, and nothing stacks. The team got a run sheet of everything scheduled, with screenshots. It went live at 12:00 AM Eastern on November 1. ### The worst bug came from Instagram's browser, and no test could catch it. The owner found it, on a phone, by accident. Instagram's in-app browser autofilled the pop-up and dragged the whole page down to the footer. For a store where more than half the traffic comes from those apps, that's about as bad as a bug gets. My tests had missed it for good reasons that were still my fault. They ran in Chromium and never in an in-app browser. They blocked Klaviyo so the pop-up wouldn't cover the page. They ran on a preview, never on the live site. And even my simulated autofill didn't move the page before the fix, so no script would have caught this one. I fixed it that day. Email fields now stay switched off until they're near the screen, and the pop-up returns you to your spot when it closes. Then I wrote a plan for finding bugs nobody has imagined yet. ### The store gets checked every morning. Every change gets reviewed. Every code change got an automated security review: 233 sessions, made up of 209 first reviews and 24 follow-ups. They flagged XSS paths, an unsafe way to push live, a discount scope error, and a discount retirement that reached too far. Some were false alarms. Several changed the code, including the push tool, the launch script and the indexing queue. A scheduled agent also runs a daily check with a report on file for every day since October 22. ### The ads sell the interactive pages by showing them working. The agency runs the ads. On October 23 I made three vertical motion ads in Remotion, written in code and rendered from it, each one selling an interactive page by showing it working. They use the store's own numbers and jokes. The creative lead and the client's team supplied the studio photos, renders and footage, and they made the show. My part was selecting, cutting, compositing and building. ### Agents let one person cover what used to take a team. This is what they ran on. Every tool I touched for this job, and what I used it for. Agencies and the client's team run some others, and those aren't on the list. --- # Dot https://davidmonson.ai/work/dot ## I created an AI character teachers crocheted. Then we took its face off. Over two years I took Dot from a first sketch to a character guide, 12 emotions and 72 poses that anyone at SchoolAI could use. Teachers made it in yarn and frosting. Then, about seven weeks after my memo arguing to keep the face, the face came out of student and teacher interactions. SchoolAI. I designed Dot and led the character from the first sketch to the redesign. 2024 to 2026. - 72: Poses, June 2025 - 0: Faces in student and teacher interactions after that launch - Yarn: Also nails, cookies, a pumpkin and a wedding bouquet - 7 weeks: About how long from my memo keeping the face to the launch without it ### What happened. Nobody crochets software. People crochet Dot. I call Dot education's most beloved AI character, with a cult following among teachers. That's my framing, and I don't have a chart for it. What I have is on this page: the system behind the character, the fan work in yarn and frosting, and the launch that took the face out of the product. ### A chat box wasn't going to earn a teacher's trust. SchoolAI puts AI in front of teachers and students every day. Trust was the whole job. I thought a character could help, if it had a real personality and stayed the same everywhere. So in August 2024 I started asking image models for one, and then I wrote down who it was. A character looks like an art decision. It isn't only that. How it looks sets how people treat it. How people treat it sets what the product is responsible for. That's what this page is really about. ### I picked a plain blue sphere because nobody could mistake it for a person. I started in DALL-E with a prompt for four fun, playful assistant characters for K-12 students. Then came rounds in Midjourney: a blueberry, an apple, a sun, a seal, a plant, a flame, inflatables and balloons. I art directed. The models made the pixels. The plain blue sphere won. A blueberry is a joke you get once. A sphere points at nothing, so it can become anything later. It's also plainly not a person, and that mattered more than I knew in 2024. ### I wrote Dot down as a system, so anyone could use it without drawing it. Art alone doesn't scale across a company. So I wrote the Dot Character Guide: who Dot is, how Dot talks, how Dot moves, and what each team had to do about it. Then came 12 emotions, and later 72 poses, so anyone on the team could use Dot without drawing it and Dot would still feel like Dot. ### A school Chromebook decides how much personality you get. Motion is where a character comes alive, and classrooms run on low-powered laptops. I tested with Kling in January 2025, then wrote the brief for a Rive motion designer. It put performance above complexity, in writing. By April we had focused, rocking, steady and hover states. In June a design contractor made the art for a live puppet in Adobe Character Animator, and colleagues performed it. The details are the character. A blink on a timer looks like a machine. So the brief asked for variation, and the character got much of its life from what it did when nobody was talking to it. ### Once Dot was a system, marketing could run with it. Office screens. A tower on an expo floor. A booth at ISTE. A billboard truck outside a conference. A live puppet for conference demos. Stickers, an iMessage sticker pack and a Dot song made with Suno. Once a character is a system, marketing can pick it up and run with it without waiting for design. ### Then people started making their own, and nobody asked them to. Nobody asked them to. Teachers crocheted Dot. Someone baked it, someone carved it, someone painted it on their nails. One ended up in a wedding bouquet. One teacher crocheted a Dot while running a crochet camp on the day we launched a new interface, then posted about it. This is as close as I get to a metric. "Kids draw Dot. Teachers put Dot stickers on laptops." I wrote that in a memo in 2026, as a reason to keep the face. ### In February 2026 I argued the risk wasn't the face. It was the behavior. Lawmakers had started looking hard at AI characters in products for students. Inside SchoolAI we were debating whether to keep Dot, change Dot or remove Dot. So I wrote down my position and said up front: "I designed Dot. I know that creates bias." I called it "a strong opinion, not a final decision." My argument: the real risk isn't the face. The harm in the news came from how AI systems behaved. First-person emotional bonding language. Role-played relationships. Encouraged dependency. Failing to get help when a student was in crisis. "If we strip Dot to a logo and change nothing about how the AI responds, we haven't made students safer. We've just made ourselves feel like we did something." The questions that matter are whether teachers can see what's happening, whether students get help, and whether the guardrails hold. I also asked people to poke holes in it. Then I proposed a system with options, so a law or a customer wouldn't force a rebrand. ### About seven weeks later, we went further: the face came out. We went further than my memo. At the start of April 2026, Dot stopped appearing with a face in student and teacher interactions. A clean abstract circle took its place, live in the product. The mascot stayed in the brand, the marketing and the community. We said why in public, in plain words. People have long formed one-sided bonds with characters on screen, and children especially. A TV character is passive. AI responds, remembers and adapts. Give it a name, a face and a steady personality and that bond gets stronger. Students should know what kind of thing they're talking to. No regulation told us to do it. And the words changed along with the art. When a student is stuck, Dot asks the next question instead of giving the answer. When a student is frustrated, Dot says what it observes: "it looks like this is frustrating." It doesn't say "I know exactly how you feel." A mascot doesn't talk back. The thing in the chat does. That's the line we ended up drawing. It's easier to see when you put what I wrote in 2024 next to what we shipped in 2026. ### I wrote the 2024 lines for warmth. The 2026 lines keep the warmth and drop the pretending. Those fan photos are mostly adults having fun, and I'm glad they exist. The harder question is what a young student does with the same character once it answers back. A crochet Dot can't do that. The one in the product can. ### A character decision is a product decision and a safety decision. Here's what I'd put in front of any team about to give an AI a personality. You're answering these whether you mean to or not. ### AI tools to explore, craft tools for the parts that had to be exact. Image models for the exploration rounds and early motion tests. Rive for the product, a live puppet for conferences, and Notion for the guide, briefs and memo. --- # SchoolAI 2.0 https://davidmonson.ai/work/schoolai ## SchoolAI grew 10x while I led design. I launched the new platform on stage. In two years SchoolAI's ARR grew 10x, past $9M, and the product reached more than a million classrooms in 200 countries. I grew design from one designer to seven, captained the redesign of the whole platform, and wrote and gave the keynote that launched it. SchoolAI. Head of Design. Strategy, product design, research, storytelling, fundraising support. August 2024 to October 2026. - 10x: ARR growth in two years - $9M+: ARR - 1M+: Classrooms - 200: Countries ### What happened. SchoolAI's ARR grew 10x in two years, to more than $9M. The product reached more than a million classrooms in 200 countries. These are company numbers, earned by a lot of people. I led design through all of it. My part: I captained the 2.0 redesign, launched it from the stage, kept going into real schools to see what we'd got wrong, led the plan that took the product to WCAG 2.1 AA, and grew a team of seven. ### Teachers kept asking for more than a chat box. SchoolAI launched Spaces in 2024: a place where a teacher could set up an AI experience for students. We visited a lot of classrooms that year, and we kept hearing the same kinds of things. "What if I could create multiple activities in one Space?" "What if I could use Spaces for more than student learning?" "Wouldn't it be cool if I could see what my students are actually thinking?" People were bending the product to fit jobs all over their schools. Answering them meant rethinking the whole platform. Everyone from engineering to sales needed to see the same picture of where we were going. ### We gave ourselves one week to draw the whole platform. In February 2025 we committed to a full design vision of 2.0 by Friday, February 14. I scoped it around five areas: Create, Organize, History, Discover and Mission Control. I wrote a brief for each, with the design challenge and what an incredible outcome would look like. We parked anything that could wait. Each parked idea got a placeholder in the information architecture, so it would have a home later. ### We missed the date, so I put an owner on every piece. By April the company needed one shared thing to look at. The goal was a full clickable prototype of 2.0 by Friday, April 18. We didn't make it. The page still says "We got close but still need to finish." I took over as captain of the finish. I split the product into areas, each with an owner and a percent done that everyone could see. I owned History, Notifications and Community, and shared Discover and Mission Control. Other designers owned the rest and did great work: Organize, Create, Agendas, navigation and the student experience. Then I wrote the demo story first and made the prototype serve it. A fifth grade teacher. Dot and a podcast. A shared collection, a student Space, Mission Control with data, Discover, then making a Lincoln Space. In June I turned it into seven narrated videos and a written storyline, so everyone in the company could get it without a meeting. ### 2.0 rested on one idea: each piece makes the others better. I kept coming back to compounding value. Dot talks to you. PowerUps do the work on screen. Agendas put activities in order. Mission Control shows what students are thinking. Discover and Organize get it all to the next teacher. ### Before launch, I held the product to a QA list of 80+ fixes. A launch on a stage has a date that won't move. In the last stretch I ran a design QA list across every area of the product, more than 80 items, each with a screenshot. All interaction animations under 400 milliseconds. Radii that nest properly. One standard scrim. Icons with defined colors. I was writing the keynote at the same time. "Is the future of learning just chat?" That was the turn in the keynote. I wrote the 15-minute product overview and gave it at the Showcase in July 2025. I opened by playing back what teachers had told us, then walked through each new piece with demos. I saved the Chrome extension for "one more thing." The audience was educators, so I kept it about their day. "ANYONE can make incredible Spaces." Then 30+ breakout sessions, where they could start building. ### After launch, I set the bar at real value in 30 to 60 seconds. A bigger platform is harder to start with. So in August 2025 I started The Great Space Race. The goal was Starter Packs for every K-12 role, about 70 of them, so a new user who picks their role sees Spaces made for their actual job. The bar I set: incredible value in 30 to 60 seconds. We built a quality rubric, with one 1-to-7 score for the Space and another for how Dot performed in it. By early September we had 73 personas and had made over 80 Spaces in person at headquarters. ### Back in schools, half the teachers we watched had never launched a Space. I speak Mandarin, so on one visit in October 2025 I sat beside a teacher and researched the platform in Chinese, in a real school. The teacher had Dot write a fourth-grade Chinese story pack: four short tales, each with a moral. You learn things that way that you can't get from a dashboard. That same day we watched more than 30 educators work, a year after their first training. I wrote it up for the team that same day, and it wasn't flattering. ### I wrote the eight-week plan that took the product to WCAG 2.1 AA. School districts need this, and it opens the door to higher ed. So I wrote and owned an eight-week plan, January 13 to March 7, 2026, that pulled in every product, design and engineering team. The goal fit in a sentence: every page and feature on the design system, fully accessible, and looking how we expect. We reached full accessibility compliance for school districts, and published a conformance report dated March 31, 2026. Engineers did most of the building here. My job was the plan, the standard and keeping everyone moving. ### One designer became seven, at a company that grew around us. I started as the only designer, at about 15 people. The company reached 105, and the design team reached seven. We set up the design system, a writing guide, regular crits and the habit of going to schools. Mobbin featured our design and design system. My team went to Config together in May 2025. If you want good taste in a fast company, you have to feed it. ### I worked with the State of Utah, as the person who designed the product. A big part of this job happened outside the product. I presented to government officials with the State of Utah, including at the Capitol in December 2025. It helps to have the person who designed the product in the room, because I can answer how it works for a teacher on a real Tuesday. ### I helped raise the Series A by telling the product story. SchoolAI raised a $25M Series A with Insight Partners in April 2025. I designed the deck, ran demos, wrote the product story and pitched investors. I did the same work on later fundraising. ### We tried every AI tool we could get, down to a live Abe Lincoln. Lincoln had been in our demos since the sprint, where students interview him for a newspaper story. In August 2025 we took it further. We hired an actor, filmed him in the office against a green screen, and used HeyGen to turn that into a live avatar you could talk to, for an experimental classroom experience. Within two days we were talking to him on a laptop and on the conference room TV, and trying him in a second setting. That was how the team worked. We used every AI building tool we could get. In one feasibility spike, the design team brought Dot to life with Supabase and the OpenAI API on our own, in 30 minutes. We tried Figma Make and synthetic user testing too. A year later that habit became Lightspeed. --- # ZenniHome https://davidmonson.ai/work/zennihome ## $20M raised for a home nobody had a word for. A crowdfunding raise that had gone flat under $1M finished at $14.0M from 2,096 investors on Wefunder. We kept raising from direct investors, to $20M in all, and two agreements with the Navajo Nation were announced at $74M. As VP of Marketing and Design I led design and marketing, ran the campaign end to end, and played a key leadership role in those agreements. ZenniHome. VP of Marketing and Design. Fundraising, decks, video, webinars, open houses, press, website, support. May 2023 to September 2024. - $20M: Raised in all, crowdfunding and then direct investors - $74M: Two agreements with the Navajo Nation - $14.0M: On Wefunder - 2,096: Investors on Wefunder ### What happened. We raised $20M in all: $14.0M from 2,096 investors on Wefunder, then direct investors after it closed. In March 2024, two agreements with the Navajo Nation were announced: $50M for housing and $24M for a new factory. You can check the Wefunder total on the campaign page, and the mid-raise number on Crowdfunding Demystified. The Navajo Nation agreements are as announced in March 2024, and KSL covered them. ### People couldn't picture it, so they couldn't fund it. ZenniHome built steel homes in a factory on the Navajo Nation, near Page, Arizona. Two models, the Denizen and the Citizen. Beds and walls that move, so a small home lives bigger. Stack them and you get an apartment building. After three years of quiet work, it still didn't fit any word people knew. Tiny home, mobile home and RV were all wrong. You can't raise money for something people can't picture, so the first job was to make it easy to picture. I started in May 2023 and led design and marketing. In practice that meant the raise, the decks, the videos, the website, the press, customer support and most things touching technology. ### I rebuilt the deck around what made people gasp on tours. My first win was rewriting the pitch deck. My notes on the draft were blunt: "Icons look like clip art." "Definitely put the transforming bed!" "Make bigger deal about JV with Navajo Nation." I rebuilt it around what made people gasp on tours. After that came a short deck, an investor print deck, tear sheets, webinar decks and a conference talk. Each one had a job. I've left the money slides out of this page. ### I ran the whole campaign, from first draft to last email. Equity crowdfunding lets everyday people buy shares in a private company. We ran ours on Wefunder, with a private phase starting in May 2023 and the last money coming in May 2024. A crowdfunding podcast covered us mid-raise at $5.8M from 868 investors. It finished at $14.0M from 2,096 investors. I wrote the page and designed it. I built the deck, recorded the videos, and ran the marketing, ads, webinars and investor updates. Wefunder's team helped with strategy, and in mid-2023 I moved the updates and posts in-house. ### A month in, the line went flat. A three-line plan took it to $14.0M. By the end of June 2023 we had $943,772 from 416 investments, and the line had almost stopped climbing. The slide said so in pink. It's the honest start of this story, and it's where the real work began. The plan fit in three lines: make good content and tell a good story, get it out at scale with a clear path to invest, then use leverage to reach further. That took us from under $1M to $14.0M. ### Every raise has a flat middle, so I kept giving people news. The shape came from working with Wefunder's team and from my own notes as we went. ### Ten ad versions, each making one of three arguments. My plan called for ten versions across carousels, video ads and single images. Each one made one of three cases. I also kept a wall of screenshots of a competitor's ads and investor posts, so I knew what people saw right before ours. ### A room that moves has to be seen, so I got in front of the camera. Investors and buyers both needed to see how the rooms changed. So I got in front of the camera with a handheld mic, walking people through a home the way I would on a tour. We also made an investor video around the showroom opening and did a full photo and video shoot in October 2023. Then I kept a list of about 35 reel ideas, each built around one thing that surprised people. Real photos kept replacing renderings wherever we could get them, because a real room answers doubts that a rendering raises. ### I animated homes that didn't exist yet. Real footage only goes so far when the building hasn't been built. So in June 2023, right as the raise went public, I started making short AI-animated videos: one home in red rock, one on a lake, one in a misty forest, and whole apartment buildings by the water and in the desert. Some went out on TikTok with a simple question: would you live here? ### Housing creators got three days, a hard hat and a helicopter. Housing influencers can tell a story you can't buy, so we worked with them directly. Some of that was interviews from my desk. The bigger move was bringing them out. In February 2024 we hosted a small group of creators for three days, from Mesa to the factory to the red rock around Page. The plan was simple. Let them see it, let them film it alone, and make sure they had a great time. The first day was mostly ZenniHome. The next was all about having fun. By April 2024, some of the best creators in our space had covered us, with 1.26 million views across videos featuring our homes. ### We closed Wefunder and kept raising, to $20M in all. Crowdfunding was one track. Larger investors need a different path, so through the raise I built a direct pipeline: accredited investors, an Opportunity Zone option, and real time inside the homes. After Wefunder closed we kept going. All in, that's the $20M. It leaked at first, and my notes say so. Serious leads were raising their hands and not closing. One reason was as small as "trouble logging into the data room." So I ran a retro and wrote the fix: one owner for the pipeline, a response in under 15 minutes with no exceptions, a daily standup, and every step of the funnel tested by hand. ### Every investor walked one straight line. We mapped the whole process on a shared board. It got big fast, with every lead source, every form field and every follow-up. Then a sticky note on the board said what it needed: "Need to show single linear happy path for prospective investor." So we drew it. Here's that path, with the real steps behind it. Tap a step to open it. ### 17 emails, each with one big piece of news. Every lead got the drip automatically. Each email carried one real story to make people want to dig deeper, and the same clear next steps: set up a call, get into the data room, sign the NDA. ### I played a key leadership role in two agreements with the Navajo Nation. The factory was in a repurposed warehouse from the old Navajo Generating Station, employing Navajo workers near where the plant had shut down. By March 2024 there were 135 of them on the payroll. The Nation's president had launched an initiative to build 1,000 homes. There was a real fit, and it took a lot of work to build. I traveled to meet the tribal council and the president multiple times. I pitched, ran meetings, hosted and toured council members, helped write the agreement, and designed, planned, ran and filmed the press conference. KSL named me as a ZenniHome spokesman in its coverage. Two agreements were announced in March 2024: $50M for housing, about 200 homes, and $24M for a new factory in Page. These are the agreements as announced at the time. ### You can't market a factory from a laptop. I walked the floor, worked with employees and filmed there. That's where the good stories were: the people building the homes, and what a crane looks like setting one. We planned a series on the women building the homes, and we posted our first business customer's install on Instagram as it happened. The floor is loud. Bring a lapel mic. ### In China, I went to ask manufacturers the hard questions. In June 2024 I spent about a week in China on behalf of the company, in boardrooms, on factory floors and over tea with manufacturers. I speak Mandarin, and my notes from that trip say it plainly: speaking some of the language goes a long way. Nobody used name cards. WeChat ran everything. Factories were open to smaller orders. I won't claim a result from this one. I'll show you what I was asking, because that's the work. ### I cut the backlog first, then automated customer support with AI. I led the support team, and we had a backlog. So we held a customer support day to find out what was really going on. Then we set two goals: cut the backlog, and build an FAQ with an AI system on top of it. I fully automated customer support with AI. I'd do it in that order again. AI answers are only as good as what your team already knows, so you start with what people keep asking and who knows the answer. ### The home needed an app, so I designed one. Every home had solar, batteries, water systems, cameras and furniture that moves. That's a lot to control from a wall of switches. So I designed a smart home app in Figma, starting from a question in my notes: how would you actually interact with a smart home? The house itself is the home screen, and you tap a room to get to what's in it. It stayed a design. The app never got built. ### At a company this size, the rest of the job was mine too. A company this size has no one else to hand things to. --- # Instawork Card https://davidmonson.ai/work/instawork-card ## We freed $15M in cash flow, then gave pros a card that pays them the moment they finish. At Instawork, one hour with accounts payable led to an invoice redesign that cut days sales outstanding from 180 to under 30 and freed $15M in cash flow. From 2021 to 2023 my team also designed the Instawork Card, new payment choices and Top Pro rewards for pros, and in testing 91% picked the card as the fastest way to get paid. Instawork. Head of Design. Invoices and payment UX, Instawork Card, Payment Settings, Top Pro. 2021 to 2023. - $15M: Freed in cash flow - 180 to <30: Days sales outstanding - 91%: Picked the Instawork Card as the fastest way to get paid, 35 responses - 4.9 of 5: Ease of the new settings, 35 responses. Today's scored 4.0, 47 responses ### What happened. Days sales outstanding fell from 180 days to under 30, which freed $15M in cash flow. None of it was a new feature. For pros, the new payment settings tested at 4.9 out of 5 for ease, and 91% picked the Instawork Card as the fastest way to get paid. The first two numbers are cash. The rest come from surveys and prototype tests with small groups, so read them as signals and not as proof at scale. ### The biggest fix started with one hour in accounts payable. Days sales outstanding measures how long money takes to arrive after you've billed for it. Ours was 180 days. That's six months of waiting for money already earned. I spent one hour with accounts payable. Then we redesigned the three things that came out of that hour: invoice notifications, banners and the payment experience. DSO fell from 180 days to under 30. That freed $15M in cash flow. None of it was a new feature. It came from sitting with the people who deal with the bills, then fixing what they see. It's why I start every job the same way. There's more on that in I take the job, and in how QR check-in came out of a stadium shift. ### The card started as a drawing in a story. In 2021 I ran a storyboard exercise with the CEO and VP Product to picture Instawork two years out. That's the North Star. One of its characters is a mom of two who works shifts. In her story, she gets paid instantly to an Instawork debit card that helps her build credit. In February 2022 we turned that into the wallet prototype above: a wallet, an earnings chart, a credit score. It was vision work, made to get a whole company picturing the same thing. It informed the final product. As the card got real, the goals got sharper. We wrote them down in a way anyone in the room could argue with. ### Nobody wants surprises with their money, so we used patterns people already trust. Going from a prototype to something that holds people's money is a lot of careful, unglamorous work. My team took it through sprints from Q3 2022 into Q1 2023. We studied how debit and banking apps solved each problem, then used those patterns on purpose. This was the wrong place to invent new ways of doing things. We took the beta to product review on August 16, 2022. The plan was to start with 20 Gold+ pros in the San Francisco Bay Area, make sure the flows worked and fix bugs before a wider release. ### Polish is not optional. That's from my own notes on this project, and it covered animation, sound and haptics. Payments was our first big push to bring the app to life. Getting a debit card you've been working toward is a moment. So activation got a small animation. A small touch, so the moment felt cared for. The engineering had to catch up. Lottie made animation straightforward, but the app couldn't do sound or haptics yet. So we asked for a reusable way to add them, so the next team wouldn't have to start over. Building the first haptics and sound became a design team goal that quarter. ### We wouldn't charge anyone for instant pay until they had a real choice. In Q4 2022 the company made a call. Instapay, which paid pros right after a shift, would cost $3 unless the money went to the Instawork Card. The goal was to move pros onto the card. That's a hard thing to do to someone's paycheck. And the page wasn't ready for it. Back then, if you added a debit card, instant pay just happened. You couldn't choose how often you got paid. You couldn't really see where your money was going. So we set a rule. We couldn't charge anyone until they had a real choice. A survey backed it up: 11% of pros preferred weekly pay and another 7% wanted to choose for themselves. In the words of one pro, "Getting paid weekly." ### A simple page, with six stories behind it. One settings page had to work for six different people, depending on their Top Pro level and whether they had the card. Most of the "tons of iterations" in my notes were about these cases, and about what happens when someone loses a status they had. ### We tested it before anyone built it, and pros found it far easier. We ran it as a prototype test in Maze in February 2023. Small groups, real tasks. Pros found the new page much easier than the old one, and they understood what the card was for. ### Top Pro gave the best pros a reason to stay the best. Top Pro is the rewards program: Bronze, Silver, Gold and Platinum, earned month by month. Better levels got earlier access to shifts, pay bonuses and instant pay. Late cancellations and missed confirmations cost you. It tied everything on this page together. The card went to Gold+ first, so working well unlocked better ways to get paid. By mid 2022 the first version had stopped doing its job. We previewed a strategy for the weekly product review on July 7, 2022, and asked whether everyone agreed we should expand it with more game mechanics. The advice we got back: open with the nightclub metaphor, to set up why the current program wasn't working. The medal in the December 2022 clip above is a small animation for our Gold pros. Our first motion in the app shipped for Top Pro and referrals. ### Referrals: show people exactly where their money is. This work began as part of a project to reduce support tickets. The demo that started it came from Fullstory Friday, where we watched real sessions together. A designer on my team led it. The fix was plain: say exactly where each referral stands, every step of the way. There was also an idea to make sharing a shift generate a QR code and a link. ### What it all added up to, and what I can't prove. Over 18 months, pros on Instawork went from 400K to 4.5M. I connect that to the North Star and the product loops that came from it, with Top Pro and referrals among them. I can't give you a number that isolates either one, and I won't pretend to. ### I led the org. Most of these screens came from my team. I led the design org. Most of these screens came from the talented people on it, and I want that to be clear. --- # Building the Team https://davidmonson.ai/work/instawork-team ## I grew Instawork's design team from 2 designers to 13, and built research from zero. From 2021 to 2023, I took Instawork from two designers and no researcher to 13 designers and 2 researchers, across continents. Along the way we started a research practice, ran a survey of pros that drew 10,000 responses, built a design system with engineering and set a bar for what counted as shipped. Instawork. Head of Design. Hiring, research, rituals, design system, community. 2021 to 2023. - 13 + 2: Designers and researchers at full size - 2 to 9: Team size in about the first year of scaling - 10,000: Responses to our State of the Union survey of pros - 300+: Duplications of our first Figma community file ### What happened. Design went from two people to 13 designers and 2 researchers. We hired the company's first UX researcher, which started a new department, and our first big survey of pros drew 10,000 responses. Our Figma community file went past 300 duplications. I'm careful with team numbers, because a team's work belongs to the team. These are the ones I'll stand behind. ### Two designers can't cover a two-sided marketplace. So I wrote a one-page plan. The company was growing fast. Pros on one side, businesses on the other, phones and web on both. Design had become a bottleneck. I wrote a one-page masterplan for what I was building. It had seven lines. ### I hired for how people think when the first answer is wrong. Hiring was the first real job, and I designed the process around one thing in particular. Our guide for candidates said so in capital letters: "SHOW ITERATIONS." The first solution is never the best one. I wanted to see the failed versions of an important screen and how someone got from there to the final result. I also asked for clear descriptions of what was their own work and what was done by others. To recruit, I outlined a deck around what a good designer would want to know: the investors, the growth, the tools we used, photos of the team, and our work-a-shift program. ### We hired our first researcher, then made research everyone's job. Instawork had no UX researcher when we started. The job post for the first one asked for the usual skills, plus the line that mattered most to me: "A willingness to book shifts on our platform and conduct research as an actual user of the platform." In a talk to the company, I said we are all researchers, and asked who had worked a shift or visited a partner and learned something new. Our goal was a shift or partner visit every month. Designers did them. I did them too, and that's its own story. The trouble with lots of research is losing it. Great notes from partner visits were getting read by a handful of people and then disappearing into a busy Slack channel. So we chose tools to keep what we learned. ### 400% ownership: if something gets missed anywhere, it's everyone's fault. A team that doubles keeps breaking things it used to do by instinct. So I wrote down how we worked and fixed the parts that hurt. The rule I cared about most was one I called 400% ownership. Product, design, engineering and QA each own 100% of what reaches users. Users don't care about our schedule. Or as I put it in our strategy document: the standard is whatever is tolerated. ### Past design systems failed because nobody owned the hard first part. So we dedicated two people to it. As of mid 2022, design and engineering each had their own set of components. Neither was complete. Even the names disagreed: an "h1" meant different things on each side. This wasn't only tidiness. A designer would add a new pattern, an engineer would assume it already existed and estimate low, and the release would slip. Teams kept reinventing the same things. A standard form screen once took more than two sprints to get right, because we had never agreed on what a form was. Earlier tries had failed because nobody was dedicated to the hard first part. So the proposal with engineering was simple: one designer and one engineer, focused, until a single source of truth existed. After that, everyone could help. ### We gave our best file away, and it showed outsiders who we were. One of our team values was to post work publicly and enrich the design community. As a team matures, it should share what it learns with other teams. We published our Design File Toolkit to the Figma community. By Q3 2022, it had 220 duplications, which I called a small but meaningful start. It went past 300. It paid off in three ways. The people who wrote it learned it deeper. The whole team used the same assets right away. And it showed outsiders what kind of team we were, which helps when you're hiring. ### We started asking how AI would change design early, when it was still messy. When Stable Diffusion came out, the team opened a FigJam board called AI/ML Playtime. Two questions at the top: what tools should we look at, and how can we use them? Someone asked an AI whether it would replace product designers, and pasted in the answer. Someone else wondered about using it for profile photos. This was early, and it was messy. Nothing on this list was a shipped result. But the team was already asking how AI would change how design gets done, and that's where I've spent my time since. Later, at SchoolAI, I built Lightspeed. ### We were remote most of the year, so when we met, we made it count. "Have fun" was on our list of team values, with the note that we'd get together several times a year to play, explore and experiment. Jokes were explicitly allowed. We were remote and async across time zones most of the year. This was the part I liked most. --- # QR Check-In https://davidmonson.ai/work/qr ## 50,000+ QR scans in 15 weeks, at a check-in so fast it backed up security. Stadium workers used to line up to sign a paper sheet. In 2022 we gave every pro a code that a captain scans in about 7 seconds, against 15 to 20 to look someone up, and kept the one thing paper got right: a person at the door. Instawork. I led design and went on shift. A designer on my team led the screens. Field research, clock-in strategy, QR codes and captain tooling. 2021 to 2022. - 50,000+: QR scans, in and out, since June 16 - 7s: Average scan with the code ready, against 15 to 20 seconds to look someone up - 0.04%: Time-theft rate for QR, week of Sep 12. A four-digit code was 0.87% - 10-17%: Clock-ins needing an adjustment with QR, late Aug to mid Sep. Codes were 22-23% ### What happened. QR launched on June 16, 2022, on every shift with a captain. By our clock-in review on September 27, it had passed 50,000 scans. It was the fastest supervised method, it needed the fewest adjustments, and its time-theft rate ran near zero. QR ran only on shifts with captains, so read these as what it did where we used it. ### Paper was still the most trusted timesheet in the business. That was the awkward fact in our own review. Instawork was a phone-first marketplace, and when money was on the line, businesses still trusted a person with a clipboard. Paper works for a reason. Someone looks at you, sees you're there, and writes it down. They also see whether you're dressed for the job and fit to work. Our four-digit codes were quick, but a code can go on a whiteboard for anyone to read, and nobody has to look at you at all. Our notes from October 2021 were blunt: some fraud, lots of tickets, and our best pros deserving the benefit of the doubt. Timesheet complaints were also getting in the way of something bigger, a future where partners prepaid. ### On shift at Allegiant, our pros walked in while other agencies waited in line. On December 3, 2021, I worked shifts at Allegiant Stadium, and I spent part of it studying check-in. QR was already being tested, months before it launched. Our pros scanned a code and walked in, and it was blazing fast. Other vendors had huge, slow lines and paper sign-in. You don't get that from a dashboard. Standing there, you can see it: the same door, the same morning, and two completely different days at work depending on which agency sent you. ### Every venue breaks it in a different way. I ran a clock-in workshop with the CEO, built on one statement: there are different types of venues with different needs, so do we agree that no single solution will work? We put our in-person research on the wall, with photos of whiteboards from real sites. In my notes: different horses for different courses. We kept it to three problems to weigh at every venue: speed, fraud prevention, and basics like connectivity and power. And we kept coming back to the things a human does at the door: last-step verification, moving people to stations, knowing who's who, and making both sides feel confident. ### Two people at that door, wanting different things. These two lines sat at the top of our board on February 4, 2022, when a Discovery Day put QR at the center. Everything got judged against them. That board also mapped everything a code could do across a shift, and one sticky said it best: a human is important for clock-in and mid-shift, but not for clock-out. ### We chose QR because it kept the human check at nearly the speed of a code. We laid every method side by side. Geofence won on speed, with no person involved. It was also the hardest to trust, and a precise fence still leaves places to hide. Looking someone up was the most human, and the slowest. QR was the only one that kept the human check at close to the speed of a code. ### The flow came down to who has to remember to do what. Two early versions from the FigJam. The difference is who has to remember to do what. In the first, arriving does the work: being near the venue readies the code for everyone. In the second, the code is simply part of the app, and people pull it up themselves. ### Many versions later, the code got simpler: your face, your level, one code. Active design started on April 6, 2022. I look for this in anyone I hire: the failed versions, and how you got from there to the answer. Here's how the code changed. An early idea put the pro's face inside the code. Later versions got simpler, with the photo above the code and the code on a plain white card on brand blue. ### The real test was the captain's scanner: a line of people, sometimes poor light, and no time. The real test was the scanner. A captain has a line of people, sometimes poor light, and no time. We went through direct-to-profile versions, a scanning view, manual adjustments, and dedicated clock-in and clock-out modes, then spent June testing how to rate someone without slowing the line. ### We tested at stadiums, launched to every captain shift, then listened. Stadiums were the initial test. QR was being tested there well before launch, which is what I saw at Allegiant. On June 16, 2022, it launched as an MVP on every shift with a captain. Then we watched it work. People on our team went on site and wrote down what they saw. Captains told us directly, in a Slack channel that grew to hundreds of testimonials. ### The charts: QR grew fast, needed the fewest corrections, and nearly ended time theft. I redrew these from the charts in our review so they'd stay sharp. Hover or tap for a week's numbers, or open the table. The last four full weeks are exact figures from the review. The earlier points were read off the original charts, so treat them as very close. ### Then we asked what else a code could do. The review ended on what might be next. Our principles for that: optimize for trust and for speed, create a great experience on both sides, and limit interruptions to the work. The big questions were honest ones. What share of partners would use QR without captains? What about small groups of 1 to 25? Wouldn't a partner scanning people go against convenience? Early conversations said no. A lot of work and worry already happens after a shift, cleaning up timesheets. Partners we talked to saw this as an upgrade. ### My team designed it. I went and looked. A designer on my team owned the QR file and led the screens. Another designer explored early versions. Our researcher was on the board, engineers built it, and captains made it work on the ground. --- # I Take the Job https://davidmonson.ai/work/field ## I do the job first. That's where a 7‑second check-in and a $15M fix came from. Before I build anything for people, I go do what they do: stadium, warehouse, hospitality and catering jobs at Instawork, a factory floor at ZenniHome, a classroom in Mandarin at SchoolAI. At Instawork, what I saw at stadium doors fed QR check-in, which passed 50,000 scans at about 7 seconds each, and one hour with accounts payable led to the fix that freed $15M in cash flow. Here's what I saw, and what changed because of it. Field research at Instawork, ZenniHome and SchoolAI. 2021 to 2025. - $15M: Freed in cash flow, after one hour with accounts payable - 50,000+: QR scans by September 2022 - 7s: Per scan, against 15 to 20 seconds to look someone up - Monthly: The team's goal for a shift or a partner visit ### What changed because I went. Stadium shifts fed QR check-in, which launched on captain shifts in June 2022 and passed 50,000 scans by September. One hour with accounts payable led to the fix that freed $15M. A mall shift made me push harder on profiles and certifications. Working shifts became a team value too, with a goal of one shift or partner visit every month. --- # Symantec https://davidmonson.ai/work/symantec ## Buried features became 2 clicks away, and one fix saved over $1M a year. A customer once asked for a feature we'd shipped months earlier. It was 14 tabs away. From 2017 to 2019 I went to see how security analysts really work, then helped take Symantec's enterprise products dark, flat and consistent, leading UX for 10+ of them. Symantec. Senior UX Designer, then UX Manager. Security Analytics, Endpoint Protection, the enterprise design system. 2017 to 2019. - 14: Horizontal tabs to find a feature that had been live for months - 2: Clicks maximum to reach any major screen, after the new navigation - $1M+: A year in support costs saved by one fix I found in user testing - 10+: Enterprise security products I led UX for ### Any major screen in 2 clicks, and $1M+ a year saved in support. A feature that had been live for months sat behind 14 horizontal tabs. After a full inventory and a flat navigation, any major screen was 2 clicks away. A fix I found in user testing saved over $1M a year in support costs. Along the way the products went dark, for analysts working in dim rooms, and moved toward one design system. Most of the work at a security company stays private, including who uses it. These are the numbers I can share. ### Powerful products that looked like the late 90s. When I arrived, Symantec had some extremely powerful products with features buried deep inside them. Security Analytics could dig through petabytes of network traffic. The people using it were sophisticated analysts protecting real assets from real threats. Our list of problems was short and hard. Data people couldn't act on. Buried features. Slow workflows. Charts that couldn't carry enough dimensions. A look that felt outdated next to the competition. ### So I went to where the software actually lives. I visited many customers and government agencies and sat with the people running their security. Most of what I saw is private, so there are no names and no site photos on this page. I can tell you what changed how I design. We paired the visits with a large survey of Security Analytics users. Here's some of what we learned. ### Customers couldn't find our features, so we made every major screen 2 clicks away. A large customer gave us a hard time for not delivering a major feature we'd promised six months earlier. The product manager said it had been live for over four months, and scrolled through 14 horizontal tabs to find it. Silence. Then the customer said it was great, and thanks. We were building great features, and our customers couldn't find them. So we did a huge inventory and built a responsive, expandable navigation that kept everything flat. No horizontal tabs. No features nested layer after layer. 2 clicks maximum to reach any major screen. ### The room was dark and our screen hurt to look at, so we went dark. A security operations center runs with the lights off and walls of monitors. Every screen in the room was dark except ours. It literally hurt to look at. That's physical pain caused by software. So we went dark. We didn't spring it on people. First came a reskin: same layout, same features, a new look. We mapped 39 screens, down to every modal and menu. That gave analysts time to get used to the look before we changed how anything worked. Then we redesigned the dashboard itself. ### One animated ring carries six dimensions, so analysts read traffic like body language. Our users wanted more than a 2D chart can hold, so we used an animated visualization called Halo. One ring carries six dimensions at once: speed, size, color, complexity, brightness and wobble. Each one could be assigned to any metric. Healthy traffic looks like a calm purple ring. When something's wrong, the ring looks anxious and irritated. You read it the way you read a person across a table, and you know where to look next. ### A whiteboard makes it cheap to be wrong. A whiteboard makes it cheap to be wrong. I like to get an idea up early, blow it up, and draw it again. One wall started as a list: MITRE ATT&CK, alerts, kill chain, sensors, live feeds, world map. Each item grew a sketch and a pile of orange questions. How many rules do users have? Can we say for sure? What can we show here? ### Three days in one room, with an agenda built to end in decisions. In January 2019 we ran a UX workshop with designers from other teams. Printouts on the table, stickies on the walls, and an agenda built to end with decisions. One list on the board asked how to bring standards into real work: a preflight list for PMs, an epic template, and data table cells that devs start from the user story. ### A rubric, so taste stopped being the argument. Design reviews can turn into contests of opinion. The first goal of the workshop's second day was a shared measure that any new design could be judged against. We picked experience attributes together and talked through why. Design spectra did the same for the look of things. Take one quality, like shape, from organic to rectilinear. Mark where we are today and where we want to be. Then we turned the standard into four levels. Every product got a row in a tracking sheet, with a status for each item: enabled, on the roadmap, divergent or not needed. Each team added its next level to its own roadmap. ### Uniting products from many acquisitions took countless meetings, and no mandate. Symantec often grew through acquisition, so products arrived with their own styles. One acquisition alone brought more than 20 products. Earlier standards had stalled. By our count, about 250 to 300 designers across the company didn't know about each other, and they kept doing the same work twice. Our 2018 vision deck made the case for one system: patterns as shared knowledge, so design time goes to big concepts. In 2019 we surveyed the product teams on framework, style and satisfaction, then walked through 20 products one at a time. One ran on MooTools. Another on Dojo. Another on Flex. It took countless meetings and presentations with stakeholders. We won PM and VP buy-in for a phased plan, merged the efforts of the enterprise design groups, and moved toward common components, patterns and one look and feel. Nobody was forced into it. Our own FAQ said product teams wouldn't be required to adopt it. It also said they might be shamed into it, with a smiley face. ### I handed developers chart colors as math, ready to paste into code. Thousands of charts were generated on the fly. They had to be accessible, and they had to look like one family at any size. A bunch of random colors wouldn't hold up. I developed a process that stepped hue and brightness by how many colors a chart needed. Then I delivered the final HEX codes to developers as JSON arrays. They were pretty stoked about that. --- # How I got here https://davidmonson.ai/work/earlier ## I went from marketer to iOS developer to VP of Mobile and Design, and never stopped building. From 2011 to 2021 I put a state senate campaign's ads in front of 208,000 targeted people, shipped two iOS apps in one year, led the bilingual team that set up an official sister-state relationship between Utah and Qinghai, and taught UX students who went on to Amazon, Microsoft and Shopify. Each job added a new way to ship. None of them took the building away. Here's how it went, and what each step taught me. Earlier work. Marketing, iOS, Flying Software Labs, Lambda School, Utah and Qinghai. 2011 to 2021. - 2.5M: Ad impressions in three months, for a state senate campaign - ¥2M: Grant from Qinghai Province, the first of its kind - 2: iOS apps shipped to the App Store in 2015 - $156,000: Starting salary of our top Lambda student, in a senior role ### Ten years, five kinds of job, and I shipped in every one. I've been the marketer, the developer, the designer, the manager and the teacher. Marketing taught me to find out fast if anyone cares. Code taught me to get software in front of real people. Teaching taught me to write down what good looks like, then score it. ### I started in marketing, where you find out fast if anyone cares. Before I wrote any code, I sold ideas. In 2014 I ran digital marketing for a state senate campaign and put 2.5 million ad impressions in front of a targeted 208,000 people in three months, at minimal cost. That same year I founded Music Heals Movement, which grew to 30,000 followers in months. Marketing came first, and I never dropped it. I've written launches and run growth at most of my jobs since. ### Then I learned to code, and shipped before sunrise. In 2015 I went through DevMountain to learn iOS development, and shipped two apps to the App Store that year. One was a tool for brain-injury rehabilitation. The photo below is from a late-night hackathon, when I was working as an iOS developer. Laptops open, cords everywhere, a can of soda on the table. An app came out of it and went to the App Store. Getting software in front of real people became the job. It still is. ### At Flying Software Labs, I went from writing the app to running the team. I joined in 2015 as an iOS developer and became VP of Mobile and Design. Flight schools have a scheduling problem most software never sees. Booking one flight means checking flight hours, duty hours, certifications, aircraft hours, maintenance records and licenses. I shipped two Swift apps for it, created the company's UI and UX function, and redesigned five aviation products across mobile and desktop. ### In fluent Mandarin, I led a bilingual team that linked Utah and Qinghai. I became fluent in Mandarin over more than three years of study and work in Taiwan and China. From 2011 to 2019 I served on the board of the Utah-Qinghai Ecopartnership, which connected the State of Utah with Qinghai province. I led a bilingual team of six across both governments to set up the official sister-state relationship. I also secured a first-of-its-kind ¥2M grant from Qinghai Province by managing technology transfer across companies and universities. In November 2016, I sat at this table signing partnership agreements between the two. ### At Lambda School, the product was people. I joined in 2019 as UX Manager. We put it plainly: our product is our students, and UX hiring managers are the customers. Students started from zero and had months to get ready for their first product team. The wall in those first weeks said what mattered: projects good enough to impress hiring managers, projects deployed to real users, and students who thrive until the day they get hired. ### It adds up to an executive who still builds. I've been the marketer, the developer, the designer, the manager and the teacher. When I run a team now, I know what each of those jobs feels like at 4:23 a.m. The rubrics came with me too. I test AI products the same way today: write down what good looks like, then score every run. That's how someone ends up an executive who still builds. I never found a good reason to stop.