Field notes
How we ship.
Playbooks, teardowns, and hard-won lessons from building AI products that reach real users.
-
Weeks, not quarters: how a small studio out-ships a big team
Most AI projects do not die from a bad model. They die in the gap between the person who understands the problem and the person writing the code.
Read → -
Why your AI feature feels bolted on
Most AI features feel bolted on because the AI got added as a checkbox instead of doing the job the user actually came to do.
Read → -
The eval you skip is why your AI ships broken
Evaluation is not a launch gate you bolt on at the end. It is how you steer while you build, and skipping it is why AI ships confidently wrong.
Read → -
What a product audit actually finds
After two dozen audits, the finding is almost never the tech. It is that nobody agreed what done means, or the data was never there.
Read → -
Building for 23 million users taught me to delete things
At scale every feature is a tax. I grew a paid product from $7 to $20 by removing confusion, not by adding capability.
Read → -
How I deliver without a fifty-person agency
The studio model, a senior lead plus a small vetted builder network, out-delivers both a lone freelancer and a big agency on the first version.
Read → -
Stop building the demo
The demo that looks done is a separate product you will rebuild for real. Build the thin real thing instead.
Read → -
RAG is not a moat
Anyone can wire the same stack in a weekend. The moat is proprietary data, an eval harness, distribution, and trust.
Read → -
How to scope an AI project so it does not balloon
Scope from the one metric that has to move and the smallest slice that proves it, then do the data reality check before anything else.
Read → -
Shipping AI where a mistake costs money
In fintech a wrong answer is not a funny screenshot, it costs someone real money. That changes what you are willing to ship.
Read →