Writing
Notes on building AI systems that survive contact with production.
Mostly things I got wrong first. Retrieval, evaluation, latency, and the data plumbing nobody puts in the demo.
5 posts shown
Legal RAG: citations before answers
In legal retrieval, fluent text is not the product. The product is traceability. This is why I inverted the usual RAG contract and made retrieval quality the gate.Read note
MRI Tumor Viewer: ship the model to the browser
The biggest gain in this viewer did not come from another point of accuracy. It came from moving inference to the reviewer’s machine and removing the upload step.Read note
PolicyFabric: make the contract graph visible
Data contracts usually fail far from the change that caused them. PolicyFabric came out of wanting the contract graph itself to be the product, not a hidden implementation detail.Read note
Your backtest is lying to you
Point-in-time correctness is the difference between a financial ML pipeline that works and one that only appears to. Here is where the leaks hide and how to design them out.Read note