Inference and edge-runtime engineer. I land narrow, tested fixes in upstream computer-vision and ML codebases, and build local-first tooling that verifies a system is still what it claims to be. Most of the work sits at the boundary where a model's output stops and an auditable decision has to be made.

1,029 contributions in the trailing year (GitHub contributions graph, 2026-09-12).

Eight merged pull requests across six third-party projects. I contribute upstream as an outside contributor — I don't maintain these projects.

Open now

• roboflow/inference#2983 — draft, open 2026-09-12: carry image dimensions on empty Florence-2 VLM results.
• huggingface/transformers#48022 — open: warn when pad_token_id is in the eos_token_id list.

I also have closed attempts and small doc/typo merges; the merged set above is the substantive record, so I don't enumerate the rest here.

Tools that answer "is this still the same thing?" — published and installable:

• trustcard — cryptographic trust infrastructure for MCP servers: signed manifests, TOFU pinning, two-gate call-time enforcement, and a scanner that probes a server and emits a scorecard instead of trusting the agent's self-attested booleans. Published to npm as mcp-trustcard (v3.1.0, 2026-09-12).
• cvconform — differential conformance for computer-vision models: compares ONNX/TensorRT exports against source checkpoints across backends and precisions to catch silent numeric drift that passes unit tests but breaks a production pipeline. Published to PyPI as cvconform (0.1.1).
• repo-archaeologist — offline, zero-dependency architecture / risk / onboarding briefing for an unfamiliar repo. Published to PyPI as repo-archaeologist (0.1.2).

• apple-quality-recognition-engine — real-time CV pipeline for apple variety detection and USDA-style grading. YOLO26 detects; a grading_policy.yaml grades. Defects bind via Intersection-of-Area, low-confidence frames are harvested as training data, and it runs on the Apple Neural Engine through CoreML. A prototype with a placeholder model — not a validated commercial grader.
• mac-ai-os — a local-first AI operating system for macOS. Its PR-governance system is an evidence-producing readiness state machine that captures real command output tied to a commit, so an agent can't fake "tests pass" with a boolean. Public with CI. Caveat: the public tree is a snapshot and trails local work.
• aafp-commons — a signed, append-only evidence ledger for agents: Ed25519-signed packets with conflict and resolution tracking, readable over MCP/CLI. Hook-and-file only; no daemon.
• X-MaC — macOS system sanitizer and discovery tool in Rust + SwiftUI. Every engine is read-only; privacy redaction is on by default, and remediation scripts ship with destructive commands commented out.

I also work at the hardware/deployment boundary: filed public issues for missing prebuilt wheels on newer NVIDIA targets — state-spaces/mamba#1019 (sm_103 / B300), Dao-AILab/causal-conv1d#119 (sm_121 / GB10).

• Reproducer first. Failure case → trace → fix → regression test. Fixes land with a scoped test, not a full-suite claim.
• Detection identifies; algorithms decide. Keep judgment out of the weights and in auditable config, so a rule change doesn't require retraining.
• Verification has to produce evidence, not consume it. A boolean from the agent that ran the action is not proof.

20 public models and 1 dataset on HuggingFace (2026-09-12), mostly small fine-tunes and GGUF/MLX conversions for local agent loops.