Systems + verification engineer working on verifiable AI and Physical AI.
I build evaluation and evidence workflows that make AI-agent and cyber-physical behavior independently inspectable: reproducible experiments, explicit verification contracts, failure analysis, provenance, robotics simulation, and requirements-to-evidence systems.
I verify AI systems — from coding agents to robots in simulation.
▶ Watch the 3:44 case study — the two runtime defects, the frozen campaign, and the 0/40 result.
A closed-loop Isaac Sim + ROS 2 Franka experiment with seeded sensor noise and causal filtering.
The current public campaign was preregistered: 20 seeds × 2 conditions = 40 scheduled runs, frozen before the first run. All 40 then executed validly — with preserved raw evidence, reproducible paired analysis, and explicit claim boundaries.
- on the three headline paired metrics — tracking RMS, disturbance attenuation, and true articulation-position RMS — filtering won in 20/20 paired seeds;
- 0/40 runs passed the full certification gauntlet;
- the acceptance thresholds were not relaxed after seeing the result;
- simulation only — no physical-hardware, safety, certification, production-readiness, or real-world-transfer claim.
Pinned case study: M4_CASE_STUDY.md @ 24bf738 — the commit-pinned evidence-backed record for these campaign claims.
That distinction is intentional: execution success and metric improvement are evidence, not permission to overclaim.
- Fieldheld Recorder — independent evidence for AI-assisted software changes.
- Agentic Systems Verifier — Case Study — sanitized verification case study.
- Repo Preflight Drift Scanner — small verification tooling for repository drift/preflight checks.
I am also working privately on fail-closed AI evaluation, verifier disagreement, quarantine, provenance, and AI-scaffolding experiments. Public artifacts will be released only after their evidence, privacy, and claim-review gates are satisfied rather than exposing private working history prematurely.
- Physical AI Research — sanitized export of the research monorepo: NVIDIA Cosmos and Alpamayo evaluation harnesses on GB10 hardware, with frozen manifests, per-file provenance and checksums.
- Articulated Robot Manipulation — Franka Panda manipulation and robotics integration in Isaac Sim.
- Robotics Ontology — robotics terminology, interfaces, and systems/MBSE-style traceability examples.
- Distributed Robotics Infrastructure — supporting edge/host/cloud robotics architecture reference.
I am most interested in the engineering layer between promising AI capability and trustworthy system use:
- AI evaluation and verification;
- agent evidence and reproducibility;
- Physical AI / robotics simulation and sim-to-real evaluation;
- AI scaffolding, harnesses, tools, and evaluation loops;
- systems architecture, MBSE, traceability, and technical decision-making;
- failure modes, explicit uncertainty, and honest non-claims.
My public repositories are personal, clean-room work. They contain no employer, customer, classified, export-controlled, proprietary, or private operational material, and no employer endorsement is implied.