AI engineer. I build software by running teams of AI coding agents — and most of my effort goes into the part that makes that trustworthy: gates, tests, and proofs that what the agents built is actually correct.
Former Marine Corps signals-intelligence team chief and Arabic instructor. M.S. Computer Science in progress at Georgia Tech.
Verified output, not vibes
| Project | What it is | Proof point |
|---|---|---|
| alchemist | C-to-Rust translation on a local LLM. Every function is run against the compiled C original; disagree on one byte and it doesn't ship. | zlib round-trips byte-exact · 667 commits |
| signet | Policy gate that checks every action an LLM agent proposes before it runs | Apache-2.0, on PyPI |
| attnroute | Context routing for Claude Code — the right files in the prompt, per turn | MIT · 30 stars |
| SPIRE | Contested-logistics operating system for the USMC. Offline, role-shaped views. | Live demo · 517 commits |
| SchoolCircle | Grounded, offline LMS: every answer cites the manual or refuses | Live app · 249 of 349 commits |
| harakat | Arabic diacritization | 2.29% DER in 6.7 MB, 62x smaller than SOTA |
| nahawi | Arabic grammatical error correction | 78.84% F0.5 on QALB-2014, 124M params |
| defense-trackers | Open defense-innovation datasets, hash-chained and reproducible | Live site |
| Naval Letter Generator | SECNAV M-5216.5 correspondence, offline on NMCI | The original behind DonDocs, now used across Marine Corps commands |
Private, available to discuss: an autonomy stack for uncrewed surface vessels, built by seven coordinated Claude Code agent teams (1,300+ merged pull requests), and a governed red-team agent.
Python · Go · Rust · TypeScript · PyTorch · Claude Code