I build the product and engineering organizations that take AI across the trust bar where errors cost bodies, not dashboards — surgical robotics through FDA, enterprise AI at scale, agentic systems shipped in the open.
Twenty years, one problem at increasing stakes: making powerful systems trustworthy enough to act on the physical world. Twelve years in real-time distributed systems (MathWorks) — the substrate agent orchestration now runs on. Multi-team embedded engineering on a Class III robotic-surgery platform through design transfer and FDA/CE authorization (Medtronic Hugo™ — IEC 62304, ISO 13485/14971), where I built the SOUP governance function from scratch: automation and intelligence replacing manual vulnerability review of COTS, OTS, and open-source components — and owned it end to end, including weekly sessions with General Counsel on EULA compliance, reselling terms, and IP review. Engineering risk and legal-regulatory risk, held in one seat. Then concept-to-commercialization leadership of new AI/analytics service offerings on a $175M IoT SaaS portfolio (Fortive) — new monetization models built on customer analytics and master data management, $1M incremental revenue in year one. Now: a 200+ person global product and engineering org at Schneider Electric, and the vision and macro-architecture for a production GenAI platform spanning ideation to engineering to post-release support.
Along the way I kept volunteering across the walls: cadaver studies with clinical-trial and human-factors engineers, market research and user surveys for long-horizon roadmaps, partnerships with marketing and business development. Engineering, clinical, commercial, and regulatory — I've worked inside all four, which is the only way I know to build systems that survive contact with all four.
I come from a family of surgeons, and I've seen up close what it costs when the right information isn't at the point of care. This work is personal.
Every repo keeps a decision record (DECISIONS.md, ADRs under docs/, or a decisions section in its README) — because architecture judgment is the job — and reports its own failures, because a pipeline that claims 100% is either being fed clean data or lying to you.
🏥 surgical-fhir-pipeline — shipped. Maps synthetic robotic-surgery telemetry to FHIR R4B behind a FHIR REST API (a scoped $everything, CapabilityStatement, OperationOutcome errors; not a full FHIR server), and refuses to hide what the crossing destroys: 80% of cases exchangeable, every lossy mapping logged, terminology bindings marked PROVISIONAL with a test that fails if anyone promotes them without the work. Ships a Provenance resource per case and a SQLite-backed governance-KPI endpoint tracking exchangeability trends across runs. Its trust-status types are isolated in their own module; whether to share them with prior-auth-agent as a library was weighed and declined on the evidence (D11). Interoperability as a governance problem, treated like one. Live demo.
🔎 device-recall-explainer — shipped. An agent that answers "why did this change?" about FDA device recalls and refuses to print any number it cannot trace to SQL. Every number is a claim citing a cell of a Zetaris query over public openFDA data; a deterministic evaluator — plain code, not the model — re-reads each cited result and rejects a draft with a named reason (no citation, wrong scope, value mismatch, an unclaimed number), sends the reasons back for a redraft on a larger Nemotron model, and prints nothing it cannot verify. A recorded run caught what a headline count hides: 54 recall records for one product code were a single event from a single firm. Gate evaluation on 18 fixed drafts: 11 of 11 bad ones rejected with the expected reason, 4 of 4 good ones accepted, 3 known gaps listed rather than hidden. H-MEM memory recalls earlier verified claims as hints, never as evidence. 348 tests, a Docker replay that needs no keys, five recorded sample runs, and a replay demo. Limits, stated up front: the default exploratory mode produced no verified answer in its recorded run (the demonstrated mode is the playbook, which covers one question shape), and the scope check is a word check, not understanding. Built for the Open Agent Hackathon 2026.
⚖️ prior-auth-agent — in progress. A LangGraph agent drafting prior-authorization determinations: FHIR bundle + CPT → RAG policy retrieval → per-criterion evidence extraction → determination. One hard rule: it never drafts a denial (schema emits only approve / insufficient_evidence; humans make adverse calls). A bilateral citation gate rejects any “met” criterion missing both a policy quote and a FHIR citation. A deterministic confidence gate — not the model — decides auto vs. human review. Backed by policy-as-code criteria beside RAG, a hash-chained audit log, a 4-dimension golden eval suite with zero-cost cassette CI, and a Terraform PHI-boundary reference architecture (Fargate/KMS/S3 object-lock; documentation-grade, never applied). Per-node cost and latency telemetry feed a model-tiering analysis, and the audit packet reports a confidence-calibration metric (Brier score). It also exposes the policy corpus as a read-only MCP server (every response carries its citation), measures retrieval quality on its own (recall@k, MRR, gated separately from determination accuracy), and runs a four-case adversarial suite against chart-embedded instructions. Live demo. Next: populate the eval baseline with a live run, calibrate the 0.85 threshold, fix vacuous auto-approve on empty criteria.
📡 fleet-anomaly-evals — early. An evaluation harness for anomaly detection on a simulated robot fleet, scored against labeled ground truth. Built so far: a seeded, physics-lite reference simulator, three fault injectors (motor thermal runaway, sensor drift, comms loss) and a fault ledger, with the key decision recorded as an ADR: ground truth is injection time, and when a fault became detectable is logged alongside it, so a detector's latency splits into signal-limited lag and algorithm lag. 65 tests. Not built yet: the bearing-wear injector (the wear curve is a deliberate open design decision), any detector, and the evaluation harness itself.
🧬 clinical-coding-lora — early. A QLoRA fine-tuning demonstration for ICD-10-CM coding on synthetic notes, built so every claim can be checked: a label set copied from a CMS file with its hash recorded, a test split that withholds two of six phrasing families from training, CPU baselines with bootstrap intervals, and an invalid-output rate kept apart from wrong-but-valid codes. Training and evaluation scripts are tested against a stub; the Phi-3-mini results are not run yet and will not be estimated.
🛡️ regulated-ai-evals — early. The trust layer extracted and generalized into a harness: requirement-to-test traceability captured while the tests run, content-hashed golden sets, structural coverage of the code that decides a claim, a calibration gate that reduces a decision-driving claim resting on an uncalibrated threshold unless a signed, dated deviation is recorded, and a sealed evidence bundle that detects tampering. Demonstrated against surgical-fhir-pipeline and prior-auth-agent, each pinned by commit. On prior-auth-agent, four safety properties verify with no threshold at all, while the 0.85 auto-finalization threshold is reduced to asserted: its evidence is one wrong approval among 13 recorded drafts, measured in-sample, with an interval as wide as the whole range. Limits, stated in its README: the labels come from the systems' own fixtures, so they show how the gate behaves, not that the cases are the right ones. 66 tests. What IEC 62304 taught me, expressed in pytest.
Stack: Python · Anthropic API (Opus, with a Haiku tiering analysis) · LangGraph · ChromaDB · MCP · FastAPI · FHIR R4B · GitHub Actions · Zetaris SQL · NVIDIA Nemotron via Meterless Nightshift · Docker. No patient data in any repository. Clinical-style projects use synthetic data; other projects use public datasets, with the source and terms named in each README.
If you're evaluating whether the AI depth is real: don't ask me — open the eval harness, the binding-status tests, the audit trail, and go as deep as you like. The stack took me months. The part that took twenty years is knowing what must never ship.
📍 Boston · LinkedIn · Building toward: product & engineering leadership in regulated AI · Open to: founder conversations, technical diligence, advisory