Senior AI/ML Engineer · AI Governance Runtime · Model Evaluation · World Models
I design and implement systems that make ML decisions inspectable: policy enforcement at runtime, evidence provenance, deterministic evaluation, and model-behavior analysis.
- AIGov Core — open-source Rust audit runtime with a Python SDK and CLI. It applies policy at evidence ingest, derives
VALID / INVALID / BLOCKEDfrom ledger state, and supports hash-chain verification and offline audit-export checks. Start with the runtime quickstart, architecture and data flow, and release verification checklist. - Compliance Engine — separate private implementation for policy-gated governance workflows; its source is not public.
These are separate engineering and research tracks, not one integrated product. The common question is whether a system’s conclusions can be traced to the evidence, policy, model, and evaluation conditions that produced them. AIGov Core addresses evidence and policy decisions; the world-model projects examine rollout error, counterfactual interventions, and disagreement between predicted and environment outcomes.
| Project | Scope and maturity |
|---|---|
| Agent Reliability Bench | Deterministic tool-trace protocol evaluation with batch reports; a reference benchmark, not a live-agent reliability certification. |
| Formal AI Verification | Rust policy predicates and SMT counterexample checks; proofs are limited to the supplied Boolean model. |
| Latent Dynamics Bench | Multi-seed ridge-linear rollout benchmark with persistence baseline and per-seed RMSE; no neural latent encoder or real-world validation. |
| Model-Based Evaluation Suite | Multi-seed predicted/actual return and transition-error report; can emit an AIGov Core evaluation event. Synthetic policies and dynamics only. |
| Causal ML Benchmarks | Stratified backdoor adjustment on synthetic data; assumptions are documented in the repository. |
| ML Systems Engineering | C++ reference GEMM and illustrative microbenchmark; not a BLAS performance claim. |
The smaller research repositories are deliberately scoped reference implementations. Their READMEs state what each example tests, how to run it, and what it does not establish; they are not presented as production-certified systems or peer-reviewed results.
- Matplotlib #32575 — merged fix for non-finite tick locations in
FixedLocator. - CPython #160980 — open contribution adding PathLike support to
fnmatch.filter()andfilterfalse(). - SciPy #26592 — open documentation contribution clarifying DIA sparse-format indexing.
- neuralforecast #1734 — open fix for zero and near-zero affine weights in RevIN denormalization.
My interests include AI governance engineering, reliable ML systems, and evaluation of learned dynamics. For publications and background, see my website.



