Journal
Production Agentic AI Essentials
Grounding AI Agents: Citations, Refusals, and Knowing What It Doesn't Know
Retrieval quality is where most RAG demos quietly fail. How this site's assistant cites its sources, verifies its own citations after streaming, and declines to answer when the corpus has nothing to say.
Guardrails for AI Agents: Building Safe and Predictable Systems
An assistant exposed to the open internet gets probed within days. What runs before the model sees a prompt and after it produces one: injection scanning, PII redaction, and output checks.
Observability for AI Agents: Understanding Decisions, Failures, and Recovery
When an agent gives a bad answer, the question is which step caused it. Tracing decisions across a multi-step graph, spotting failures, and the token budgets and circuit breakers that let it recover on its own.
Evaluating AI Agents: Golden Sets, Adversarial Tests, and Catching Regressions in CI
Every prompt tweak is a silent regression risk. Building a golden question set, running an adversarial subset on every pull request and the full set nightly, and scoring answers so improvements are measured rather than felt.
Engineering Reliable Retrieval: Hybrid Search, Fusion, and Source Diversity
An answer can only be as good as the context behind it. Fusing vector and full-text search into one ranking, and keeping one verbose document from crowding every other source out of the results.