TECHNICAL WRITING
Learning in the open.
Notes on building reliable AI systems and the engineering decisions behind them.
Evaluate RAG at the Retrieval Boundary
Separate missing evidence from unsupported answers using relevance labels, retrieval metrics, and a small, reviewable evaluation set.
· ~2 min read
Hybrid Search: Give Exact Matches and Meaning a Vote
Why lexical retrieval still matters, how reciprocal rank fusion combines results, and what to measure before adding a reranker.
· ~2 min read
A Larger Context Window Is Not an Evidence Strategy
What long-context research tells us about evidence placement, and how to budget context without turning every request into a document dump.
· ~2 min read
Tool-Calling Agents Need Explicit Execution Boundaries
Design tools as contracts: validate arguments, enforce authorization, bound retries, and inspect outcomes outside the language model.
· ~2 min read