the archive
Practical AI: agents, coding assistants, RAG, and evals. Applied, not hype.
A teardown of the five places retrieval quietly breaks between demo and production — chunking, stale indexes, near-duplicate context, recall vs. answer quality, and 'fine on 10 questions' — each with the fix.
read the postTDD's failing-test-first loop, applied to AI: write the eval before the prompt, let the score say when an LLM feature actually works, and stop shipping changes you can't measure.
How to get an LLM to return data you can actually use: schema-first instead of regex, constrained decoding, validation at the boundary with Zod, retry-with-error-feedback, and when a clean second pass beats one more retry.
An opinionated, no-hype tour of the open-source models, servers, and tools we'd reach for to ship an AI product in 2026 — and the five things that actually changed this year.