Built to multiply output. Built with AI, tested for real.AI-built, eval-proven.

I'm a product manager who builds with AI. I take messy, complicated problems, make them simple, build the fix, and test it so the results are real, not just a nice demo.I build solutions that multiply what a team can ship, lift performance, and solve the customer problem in front of them. Every build is measured with evals and documented in public, so the results are honest and the work moves faster.
- Focus
- AIs that do tasks for you, not just answer questions. that do real work, tested before they shipAI-built products, agents, evals
- Now building
- ShelfReady · Episode 07
Case studies
1 shipped, 4 in the queue. Real numbers onlyHow many products an AI shopping assistant can fully understand. Before: the store as it was. After: 105 fixes, each one checked by a person.Share of products an AI shopping agent can read fully. Before: the original catalog. After: 105 fixes, each approved by a person.
ShelfReady
Build 01 · Live- Problem
- More people now ask an AI assistant to shop for them. When a product's details are vague or missing, the assistant skips it, and the store never knows why.AI shopping agents skip products when catalog data is vague or missing, and merchants can't see the gap.
- Bet
- Check every product the way an AI shopper would, fix what's missing, and have a person approve each change. Then test real AI shopping before and after.Audit every product for agents, fix it with a human approving each change, then measure real agent shopping before and after.
- Result
- Almost every product is now fully readable by an AI shopper: 99.3%, up from 78%. In 360 test shopping trips, the AI never bought the wrong item.105 approved fixes took agent-ready products from 78% to 99.3%. In a 360-session shopping test across both catalogs, the AI agent never bought the wrong product.
Red Pen
Build 02 · Planned- Problem
- AI can write support replies that sound great but get facts wrong. Nobody knows how often.AI-drafted support replies read well and still get facts wrong, and nobody can say how often.
- Bet
- Write the test first: practice questions with known answers. Then improve the AI one change at a time and keep only what raises the score.Build the eval set first, then change one prompt at a time and keep only what moves the score.
- Result
- Planned: a score for each version of the AI, and a bar it must pass before it goes live.Planned: pass rate per prompt version against a ship gate.
Roadmap scorer
Build 03 · Planned- Problem
- Every request feels urgent, but the team can only do so much.Every request feels urgent and there is one team's worth of capacity.
- Bet
- Score each request by how much it helps for each week of work, and draw a clear line where the team runs out of time.Score each request by expected gain per engineering week and draw the capacity line in public.
- Result
- Planned: how many requests came in and how many the team took on. Sample requests will be labeled as samples.Planned: requests in vs commitments out. Example requests will be labeled as examples.
Restaurant allergen assistant
Build 04 · Planned- Problem
- Staff answer allergy questions from menus, recipes and memory, and those don't always match.Allergen answers come from menus, recipes and memory, and those sources don't always agree.
- Bet
- Answer only from the restaurant's own information. If two sources disagree, say "ask the kitchen" instead of guessing.Answer only from the restaurant's own sources, and send the guest to the kitchen whenever they conflict.
- Result
- Planned: how often it answers correctly, and how often it says 'ask the kitchen' when it should.
Model router
Build 05 · Planned- Problem
- Using the biggest, priciest AI for every small task costs too much.Sending every request to the largest model makes each task too expensive to price.
- Bet
- Send easy questions to a smaller, cheaper AI and hard ones to the big one. Then use Red Pen's A test that scores how often the AI gets it right. to make sure quality didn't drop.Send easy requests to a small model, escalate hard ones, and score both with Red Pen's eval set.
- Result
- Planned: the cost of each task before and after, and whether the answers got any worse.Planned: cost per task before and after, and the quality difference in points.
I use AI to build what a team couldn't build alone, then prove it works with a realI use AI to build what a team couldn't ship alone, then prove it with an testeval, not a demo.
How I run a product. Pick a step
Frame
Write the problem on one page: who is stuck, what it costs them, and which number should change.Write the problem on one page: who is stuck, what it costs them, and which number should move.
Artifact: a one-page problem summaryone-page problem briefAbout
I've led products at Visual Comfort & Co., Zillow, Spreetail and Arcadis. Now I also build the AI myself: I put it to work in real systems, measure how well it does, and write down what went wrong.Product leader at Visual Comfort & Co., Zillow, Spreetail and Arcadis. Now I build the AI myself: ship it inside the real system, measure it, and write down what broke.
More about me

Proof
Selected results from my product roles- 6 wks → 45 minvalidation and approval workflows, after leading a custom-trained LLM catalog audit agent across 1M+ SKUs and 20 teams that cleared a 150K+ SKU error backlog to zero.Checking and approving product data took six weeks. With an AI checker I led, it takes 45 minutes.
- 15–20xmonthly SKU corrections (1,000 by hand to 15–20K) by running the audit agent in production, with human-audited accuracy up from 45% to 97%.The team now fixes 15 to 20 times more product errors each month, and the fixes are right far more often.
- ~11xmonthly usage of an AI chatbot and MCP platform that acts for authenticated users, from ~200 conversations a week at launch to 10,000+ a month.An AI help chat I led grew from about 200 conversations a week to more than 10,000 a month.
- 71% → 93%decision accuracy for an AI refund agent, from an eval framework built on 80 real support tickets; it launched with zero policy-cap violations.An AI that decides on refunds went from right 71% of the time to 93%, because I built a way to test it before launch.
- $6.8Mincremental revenue from an experimentation and A/B testing framework (50+ experiments), growing paid members 29%.Testing website changes on real users before rolling them out brought in $6.8 million in extra revenue.
- $42.2Mincremental transaction value over four quarters from multivariate tests, growing active agent adoption 31% (300.2k to 400.2k).Testing page changes with real users added $42.2 million in transactions over a year.
- $4M+ ARRfrom AI-powered scheduling optimization for ShowingTime+, growing monthly tour volume from 740k to ~949k.Smarter scheduling for home tours brought in over $4 million a year in new revenue.
- 3xrequests on a centralized AI service layer powering recommendations, dynamic pricing and support automation (20,000 to 60,000 a week in six months).A shared AI system I led went from 20,000 to 60,000 requests a week in six months.
- $400M+logistics operation across 7 distribution centers with AI-powered fulfillment; vendor onboarding cut from 8 weeks to ~5.2.I helped run shipping and warehouses for a $400-million-plus operation, and cut the time to bring on a new supplier from 8 weeks to about 5.
- −31%monthly support tickets (15,000 to 10,350) from self-service documentation, with customer satisfaction up from 68 to 87.Better help pages meant 31% fewer support tickets, and happier customers.
- $1.8Mdirect procurement savings from renegotiated vendor agreements ($11.2M to $9.4M).Renegotiating supplier deals saved $1.8 million.
- −10%costs on $31M in financials managed and forecast, at 100% alignment with budget goals.I helped manage $31 million in budgets and cut costs by 10%.
Index
Builds, done and planned- 01ShelfReadyStore AI shoppers can readAgent-ready storefront99.3% AI-ready99.3% agent-ready / up from 78% after 105 approved fixes
- 02Red PenSupport replies, testedEval harnessPlanned
- 03Roadmap scorerDecide what to build nextPrioritization toolPlanned
- 04Restaurant allergen assistantSafe allergy answersGrounded answersPlanned
- 05Model routerCheaper AI, same qualityCost and qualityPlanned
- NextActivation flowSign-up testOnboarding + analyticsPlanned
- NextWeekly digestWeekly numbers noteAnalytics agentPlanned
- NextInterview synthesiserInterview summariesResearch workflowPlanned
- NextOpenAPI → Agent ToolkitConnect AI to softwareAgent toolingPlanned
- EarlierWherewiseRelocation recommenderLive soon
Off the clock
Running, Sundays and a lot of patience- Running
Since third grade
I fell for running in third grade, on the track. I kept going through grade school, high school and college, then half marathons. It's not a hobby anymore; it's just how I live.
- New York Giants
—Days since the Giants' last Super Bowl
Still counting. Still watching every Sunday.
- Tottenham Hotspur
17 years, then Bilbao
Spurs taught me patience: 17 years between trophies, and I kept showing up. Then Bilbao: Europa League champions, May 21, 2025. Same in product: keep shipping through the hard seasons.