Latest work
- ClearbellOct 28, 2026
A night-shift check for care homes.
Won prizes from You.com, Parasail and CrewAI at Hack With YOU @ AWS Builder Loft: Halloween Edition, San Francisco. - FixtureFitOct 25, 2026
Gemini 4 Argon designs and repairs manufacturing fixtures; an independent geometry checker decides whether the operator's tool can reach the part.
Repaired designs passed the independent geometry checker in 18 of 18 runs (6 sample jobs on one fixture family, each run three ways). Built at the Gemini 4 Argon SF Hackathon, Cerebral Valley × Google DeepMind. - HoldfastOct 24, 2026
Negotiates SaaS renewals by email and waits for your phone before paying.
A durable agent on Omnara. Built at Omnara's Agent Builders Hackathon. - HeatCrewOct 19, 2026
Seven AI agents turn tomorrow's forecast and California's heat rule into an approved, bilingual crew plan.
For field-ops managers of small outdoor crews with no safety staff. HeatCrew reads the National Weather Service hourly forecast and Cal/OSHA §3395, moves heavy work into the cool hours without deleting a job, and hands the manager a plan to approve and a briefing the crew hears in English and Spanish. A Safety Governor refuses “ignore the heat” and hands the override to a person. The optimizer runs as code in a sandbox, and also on RocketRide Cloud, where its plan hash matched; it now runs as an app in RocketRide's App Builder. Built at the Universal Hackathon, AWS Builder Loft. Demo crew; the home page run is a bundled sample; payments in Stripe TEST mode; messages go only to workers who opted in. - SteerProofOct 13, 2026
A release permit for refund agents: NOT AUTHORIZED until each rule is proven.
A default-deny gate sits inside the money tool, and the permit stays NOT AUTHORIZED until each rule is proven on the agent's own tickets. Shown at AWS Community Day – AI Edition. Built on Strands Agents. TEST money, fictional company. Gemini served every model call; the Bedrock adapter is wired, with 0 successful calls. - QueueTillerOct 12, 2026The inference contract controller: it enforces a customer-support promise on its own GPU.
- BreakglassOct 9, 2026The safe thing an agent may do before anyone can patch.
- Canary HeistSep 29, 2026Learn AI security by attacking an AI agent.
- BountyloopSep 28, 2026Proven-bug payouts: an agent-run vulnerability disclosure desk for seed-stage SaaS.
- StayCoveredSep 26, 2026The renewal agent that keeps eligible people on Medicaid, and is paid only for members it can prove it kept.
- SkylockMay 3, 2026hardware
A backpackable edge node for drone detection: each sighting verified, signed and replayable offline.
A backpackable NVIDIA Jetson Orin Nano with a pan-tilt camera runs a YOLO drone-vs-bird model (0.919 mAP@50 on a frozen holdout) as a TensorRT FP16 engine at 6.22 ms per frame, offline. Each detection is Ed25519-signed and hash-chained; the model's SHA-256 is checked against a signed passport. It never engages a drone. Built for the National Security Hackathon (by Army xTech), Cerebral Valley. Slack, Teams and phone-call alerts are mocked; sample receipts are synthetic; multi-node mesh is not built.
Builder across AI, ML and agent evaluation.
I build practical AI systems centered on reliability, evaluation, and enterprise deployment. I hold a PhD (UK) in Microbial Molecular Genetics from the University of Birmingham, fully funded by the Darwin Trust of Edinburgh, with 4 publications. Certified by AWS and Elastic; based in the San Francisco Bay Area. I also build on edge hardware: NVIDIA Jetson nodes with cameras on pan-tilt mounts, running models on the device. Another system I build is MemoryOS: memory governance for AI agents. It decides what an agent may remember, retrieve and act on, with 1,400+ tests and 6 framework adapters. I also built LexFusion: an AI legal-intelligence platform, presented at Stanford Law School (Stanford CodeX).