EngineeringAIEvals are just tests with worse error messagesA practical framework for building LLM evals that catch regressions you'd otherwise ship.SVSachin VermaFeb 10, 2026 · 9 min readPart of a collectionAgents from zero2 of 3The framing# Stop treating evals as research. Treat them as integration tests with fuzzy assertions. A good eval suite is: Deterministic where it can be Has clear pass/fail per case Runs in CI Owns its own dataset, versioned Pass/fail, not score-only# Scores drift. Booleans don't. ← PreviousWriting agents that don't hallucinate toolsNext →Tiny LLMs on the edge: a weekend experimentRelatedEngineeringMay 10, 20269 minApex Governor Limits in 2026: what still bitesFive years of platform updates later, the governor limit cheat-sheet you actually need at 2am during a release.EngineeringMay 2, 202612 minWriting agents that don't hallucinate toolsA small set of guardrails that cut tool-call errors by 80% on our internal stack — no fine-tuning required.MiscApr 18, 20266 minTiny LLMs on the edge: a weekend experimentRunning a 3B model on a Raspberry Pi 5 to summarize my RSS feed. Mostly it worked. Sometimes hilariously didn't.In this collectionAgents from zero01Writing agents that don't hallucinate tools02Evals are just tests with worse error messages03Tiny LLMs on the edge: a weekend experiment