Writing / Engineering journal

Notes from systems I actually build.

Architecture decisions, failure modes, evaluation, backend systems, applied AI, and lessons from product work.

Reliable RAG Begins Where the Demo Ends

A retrieval system earns trust by knowing when evidence is insufficient, not by producing an answer for every prompt.

Stable aliases, explicit policy, observability, and testable fallbacks turn model calls into infrastructure.

An agent is a stateful workflow with tools and side effects. Design it so each decision can be inspected and bounded.

Detection is a frame-level capability. Traffic reasoning needs temporal evidence, stable identities, and conservative claims.