Lead AI Engineer — secure enterprise AI in regulated healthcare.
I build production AI systems and the instrumentation that proves — or disproves — their value. Nine years shipping technical systems into organizations that had to adopt them; the last two building production AI.
Nine small, complete harnesses with one thesis: nothing ships until it passes a gate — and the gate itself must be earned.
| repo | the gate | the claim it protects |
|---|---|---|
| kappa-gate | an LLM-as-judge must clear Cohen's κ ≥ 0.70 against hand-authored labels before it is trusted | "our eval says this is good" |
| roi-gate | an ROI figure is reportable only if the model was biased against the claim — CI refuses the vendor-deck version | "this deployment is worth $X/yr" |
| phi-gate | a redaction layer must clear recall ≥ 0.95 on a labeled corpus before it may front an LLM | "the pipeline is safe for regulated data" |
| trade-gate | no order executes while the local book and the broker's snapshot disagree | "the bot knows what it holds" |
| perm-gate | zero leaks under scoped credentials — and the prompt-layer failure must stay demonstrable | "the guardrail prompt makes it safe" |
| rag-gate | an index serves only above recall@3 ≥ 0.90 on labeled queries — it already caught its own baseline's defect | "the RAG system found the right context" |
| target-gate | no outbound list is sent until identifiers, freshness, dedupe, and coverage all clear — delivery is wired behind the gate | "sales can trust this list" |
| brand-gate | a Claude Code skill brands every deliverable by default — and no file ships until brand_check.py clears it by rule number |
"this document is on brand" |
| mcp-gate | an MCP file server refuses any path outside its authorized roots — .., absolute, symlink, and %2e-encoded escapes all normalized before the check; the prompt-only version must still leak in CI |
"the model can only touch what we said" |
| repo | the gate | the claim it protects |
|---|---|---|
| intelligent-integrations-hub | getaiintegrations.com must build, and every repo it links to must exist, before a push lands | "the portfolio links to real, public work" |
Each repo ships a RESULTS.md of captured command output — regenerated by
script, never hand-edited. Each one is runnable in three commands, tested,
CI-checked, and honest about
its limits — every dataset is synthetic, and each README states what the tool
cannot do alongside what it can.
| repo | the gate | the claim it protects |
|---|---|---|
| identity-gate | an autonomous self-study agent whose performed introspection (a JSON field) is never trusted as reasoning — only the API's thinking trace counts as telemetry, ...logged append-only and kept strictly separate — rendered on a live map of every path the agent walks" |
The live map republishes its last-walked path, declared form, and confidence from sanitized telemetry every 15 minutes — shape, never content.
A continuously running agent on an isolated Linux account with no goal, no personality, and no knowledge of its own infrastructure — communicating over WhatsApp, with an optional Raspberry Pi + LCD as physical form. Loosely inspired by I Gave an AI a Body by Cyrus Clarke (Tangible Media Group, MIT Media Lab).
Small, complete browser experiments built to be played, with no business claim attached.
| repo | the game | the hook |
|---|---|---|
| gorilla-protocol | a from-scratch, content-first Unreal Engine 5.8 rebuild targeting one authored ten-minute spy-FPS vertical slice; the rejected procedural build and the Java/WebGL iterations remain preserved in the same Git history, while the paid GPU host stays stopped until local quality approval | an original Italian-speaking gorilla operative with physical-comedy mechanics, tactical encounters, and a deliberately silly banana obsession |
| monkey-no-fly-zone | a playable 3D survival flight game with streamed cities, fighter formations, homing missiles, return fire, and desktop and touch controls | a mechanically winged monkey flying head-first through escalating military airspace |
Monkey No-Fly Zone began as a small 2D arcade experiment and was rebuilt with a master 3D design audit plus focused AI sub-agents for the player rig, city streaming, combat, effects, audio, mobile UI, and release QA. The current release bundles seven modules into 621.1 KB, verifies 29 DOM hooks and six automated tests, and ships as a self-contained 4.83 MB HTML game.