Notes
Findings and half-formed thoughts on AI, product, and what actually works. Written while building.
Bookbind. The hard part was the spine width.
I have been building something called Bookbind. You type one sentence about the book you have been meaning to write, and it hands back the files a printer will actually accept....
[read]
We built Flint. The pricing decided the architecture.
A friend and I built a voice app. It is called Flint, it has been on the App Store since April, and it is on Android now. You talk, it writes. That is the entire product. It is...
[read]
Chat is a hedge, not an interface
Every AI product I have worked on began with a text box. Not because anyone designed a text box, but because at the start you do not know what people will ask for, and a text box...
[read]
Agentic RAG is a routing problem, not an intelligence problem
Agents are being pointed at questions a keyword search would answer in 30ms. A note on routing by question shape, and why trajectory evaluation is the gap nobody has closed.
[read]
The PM skill that got scarce this year
When anyone can build anything in an afternoon, the rare skill is not execution. It is being able to say what actually changed for the user, and prove it.
[read]
Evals are user research at machine speed
Most teams treat evals as a QA gate. Treat them as discovery instead. Forty labelled examples told me more about our AI product than a month of user calls.
[read]
The fourth column of every roadmap
Now, Next, Later, and the one nobody labels. On treating the roadmap as a communication artifact instead of a plan, and saying out loud which parts are guesses.
[read]
I stopped trusting vector search on its own
Dense embeddings could not find our own product codes. Hybrid retrieval took recall from 61% to 89% on our internal set. Why vector-only is a bad default in 2026.
[read]
Positioning is an argument your team has not had yet
If five people at your company give five different answers to who the product is for, that is not a messaging problem. Here is the exercise I run instead of a workshop.
[read]
Your RAG demo works. That is the problem.
A retrieval demo takes an afternoon. A retrieval product takes a quarter. Notes on the nine places I have shipped bugs between the question and the answer.
[read]
Building got fast. Learning did not.
AI collapsed the cost of shipping and did nothing to the cost of finding out you shipped the wrong thing. Why discovery is now the bottleneck, and what I changed.
[read]
