Notes · a repo a week
One open-source repository a week, and the decision behind it: what it takes off a team's plate, what it costs to own, what it puts at risk, and whether to adopt, build or buy. Written for the people who fund and run AI products, and AI here means the whole field: tabular models, forecasting and optimisation as much as language models. I run each one myself before I write about it.
Turns a stated risk appetite into the exact point where a model stops deciding and a person takes over, with a statistical guarantee, and warns when that line no longer holds.
Vector, keyword and hybrid search in a folder of files, with every version of the index kept, so retrieval ships without a new server to run and a bad re-index can be rolled back like bad code.
Standard traces for every model call your product makes, with tokens, latency and outcome per step, so the question "what does this feature cost per customer" gets an answer from evidence, and the failures that look like success get caught.
Multi-step AI jobs that survive a crash without paying twice for the steps that already worked, and retry the flaky ones on their own, on the Postgres you already run.
Simulated customers who argue, change their minds and push back, run against your agent before every release, so the failures you did not think to script show up in a test report before they show up in a complaint.
A written, tested agreement between the team that changes the data and the team whose model depends on it, so a quiet schema change stops at the build instead of reaching a customer.