Field notes on production AI, agent runtimes, evaluation, retrieval, and ML systems. Written from implementation experience, with the trade-offs left in.
AI demos are easy; AI products are hard. The gap is a runtime — the shared execution layer that handles model routing, tool authorization, long-context state, memory, and fallback. Here is what it takes to run one at scale.
Compliance answers are only trustworthy if every claim cites the right evidence. Here is how hybrid retrieval, reranking, control-aware chunking, and citation-span ranking took evidence F1@10 from 86% to 97%.
A model per document type does not scale. Here is how a single layout-aware extraction architecture replaced 23+ document-specific models, cut OCR spend 70%, and made new document types cheap to add.
No posts match this filter.