Four Decisions I Made Before Writing Any Features
Four architecture decisions that kept an enterprise expense system auditable: claim boundaries, immutable history, identity, and concurrent writes.
Read articleWriting
Writing on AI and software systems.
Articles on engineering decisions, implementation details, and the tradeoffs behind reliable production systems.
All writing
8 articles
Your Agent Still Has the Tool You Revoked
An agent can retain a revoked tool. Why long-running workflows create a TOCTOU authorization bug — and why the decisive guard belongs immediately before provider I/O.
Read articleA Timeout Doesn't Tell You Whether the Write Happened
Retrying an external write is a guess about which of three worlds you are in. Idempotency keys, business-key deduplication, and the database invariants that make the guess unnecessary.
Read articleTracing a Request You're Not Allowed to Log
Correlating one job across TypeScript, a queue, and Python — without turning routine telemetry into a second copy of the sensitive payloads it describes.
Read articleDon't Put a Model Where an If-Statement Will Do
Every model call in a routing path converts a fact into a guess. What I learned about spending probabilistic routing only where the ambiguity is real.
Read articleGenerated Charts Are Untrusted Code
Why a model-generated Mermaid diagram or ECharts spec is a program, not data, and the two-tier validation and bounded repair pipeline I built around it.
Read articleThe Model Wasn't the Hard Part: My Learnings After Shipping Multi-Agent Systems
Two and a half years of building agent systems, and what I got wrong about architecture, context, evaluations, model routing, permissions, and recovery.
Read articleRetrying an AI Agent Is Not the Same as Retrying an API Call
How leases, checkpoints, ownership-safe writes, and tool-history repair made Research-AI recoverable — and what the design still does not guarantee.
Read article