Writing · 2026

Notes on system design.

Agentic architecture, governed AI, and production ML, written from the systems I build.

04 | 3 Oct 20267 min readYour AI Eval Suite Can Become Technical DebtTurning every production failure into a permanent eval eventually makes the suite the bottleneck. Triage first: deduplicate the known, protect the novel, tier what runs where, and evaluate the triage too.

03 | 2 Oct 20265 min readA System Prompt Is Not a Security BoundaryAgents can now query databases, call APIs and run code. What limits them should live in identity, policy, sandboxing and tool authorization, not in an instruction to the model.

02 | 2 Oct 20266 min readYour Agent's Most Valuable Dataset Is Its Production FailuresCoreWeave Forge frames AI work as a run-observe-curate-improve-evaluate loop. The engineering lesson: map each failure type to an eval, test every layer of an agent, and gate every fix.

01 | 30 Sept 20266 min readTyped Decisions, Not Paragraphs: What Jev Gets Right About Agentic SystemsWhy decision steps in agent systems need typed outputs, calibrated confidence, and a human exit, and how TypeSafe AI's Jev model approaches that from the model side.