Reranking - A reranker cannot fix what retrieval missed
A cross-encoder over the 20 candidates stage 1 hands on, where gradient-boosted rankers fit, and the ceiling retrieval sets.
Articles
Thoughts on things I discover in the world of AI. I write about my experiences, learnings, and opinions.
Retrieval 2 - Fine-tuning a small embedding model gives surprising results
Eight minutes of training takes a 110M model past an 8B one, and the counterexamples you pick are the whole gain.
Retrieval 1 - Building benchmarks and why keyword search still wins
A fair test over 482,105 Amazon products, and the baselines that turn out to be harder to beat than they look.
Evaluating AI agents
Research output has no right answer to check against. How we built a judge we could trust, and what it caught.
Building AI Research Agents
A loop, the tools around it, and everything we did to keep the context from filling up.
Radiology reading times
A metric said radiologists read for 25 hours a day. Finding out why changed how we measured the work.