Director and GM at Iyara Labs, an AI-native digital agency working out of Dubai and San Francisco. Before that, email and AI consulting. MS in Marketing Analytics from Purdue.

Most of what I build sits between a business process and a model: agent harnesses, event- driven workers, and the unglamorous infrastructure that decides whether either one is actually reliable.

skill-routing-eval — An A/B harness for measuring whether the context you inject into every agent session changes what the agent does. Runs both arms, because a score from the ON arm alone can't distinguish "the map worked" from "the model didn't need it". On my own setup: 15/15 with the injection, 9/15 without. The more useful finding was the variance — three consecutive runs of an unchanged map scored 13, 15 and 11, which is why it samples repeatedly and collapses by majority instead of trusting a single pass.

claude-code-hooks — Three zero-dependency hooks: one blocks credential-shaped strings from reaching source files, one refuses to end a session that changed config without executing it, one suggests compacting at a clean boundary. The interesting part is which way each fails. The secrets guard fails open and never fires on .env, because a guard that blocks the correct answer gets disabled within a day. The verification gate fails closed, because a silently skipped gate is a broken config you find tomorrow.

windows-arm64-dev-fixes — Eight JavaScript toolchain failures on Windows 11 ARM64, each with the verbatim symptom, the actual cause, and the fix. Most share one root cause — a package ships no win32-arm64 build — surfacing far from that fact as a spawn EFTYPE, a zero-byte binary, or an npm install that exits 0 and installs nothing. All hit and resolved on a real machine.

codex-usage — A VS Code status bar reading live Codex rate limits over JSON-RPC. It started out parsing local session logs, which was wrong in a way that looked right: the newest local record is usually stale, so both windows rendered as resetting now. Now it calls the documented live endpoint and keeps the logs only as a fallback. The fiddly part is WSL, where Codex uses a Linux account store while the extension host is still Windows — launch the native binary there and you read a different store with an invalid token.

mailchimp-bounce-monitor-worker — Cloudflare Worker handling Mailchimp unsubscribe, bounce, abuse and soft-bounce webhooks. Deliverability degrades quietly, so the work is in the retry and idempotency paths rather than the happy one.

Agent evaluation, mostly. It is easy to make an agent setup feel better and hard to show that it is, and the gap between those two is where most of the effort in this space is currently going to waste.

Smaller things: claude-session-dot, a sidebar telling you which of several running agent sessions has actually stopped and is waiting on you — a hook writes the state, the extension only renders it.

Chess takes the rest of the time — 2264 rapid on Lichess, Arena International Master. deepgamecoaching.com is the coaching practice, and chess-app is the tool behind it: PGN in, style diagnosis and a training plan out, with optional Stockfish review in a web worker. It runs entirely on-device — local parsing, IndexedDB storage, no paid API anywhere in it.