I build secure enterprise systems and the AI workflows that run them.

Compliance answers are only useful if you can check them. Warrant embeds the full NIST SP 800-53 catalog — 1,014 controls from the OSCAL source — into Postgres with pgvector, retrieves the controls that bear on a question, and tags every sentence of the answer to the control ID that supports it, with the clause itself one click away. I chose the embedding model by measurement rather than habit: four candidates run on CPU against a 42-question golden set, with recall@5 and latency recorded for each. Recorded answer fixtures mean the whole pipeline replays deterministically, so the tests never need an API key.

A ground control station for simulated uncrewed aircraft, built around one rule: the console says exactly what it last heard and never pretends otherwise. The MAVLink v2 codec is hand-written and verified byte-for-byte against pymavlink vectors; the aircraft flies in its own process and speaks UDP; an ASP.NET Core service ingests into a bounded twelve-vehicle store and streams it to a React and MapLibre console over Server-Sent Events. Nothing is interpolated — at three seconds without a heartbeat a marker goes stale with its age on screen, and at fifteen it is marked lost. The requirements table names the two items still unverified rather than smoothing them over.

I built Sentinel end to end — the firmware on the device, the cloud pipeline that ingests its telemetry, and the dashboard operators watch it on. These pumps sit underground across a utility’s service area, and when one fails quietly someone’s basement floods. Sentinel’s whole job is to notice first: it watches electrical signatures for the shape of a failing pump, escalates through email and SMS until a human acknowledges, and flags devices that have simply gone silent.

An engine that turns a slash command in Google Chat into an agent run. Each workflow is a self-contained module that declares its command, its access policy and a factory returning its own agent — so adding a capability means dropping in a file, never touching the dispatcher. The engine itself stays deliberately thin: resolve the command, build the agent, run it, post back to the thread. Around that sit the parts an internal tool actually needs — per-workflow access rules, sessions that survive a multi-turn conversation, and background dispatch so nothing dies against Chat’s 30-second webhook timeout. And because these workflows read someone’s real inbox, context and actions run under two separate identities: the half that can read has no ability to write anywhere.