I work on computer vision, inference reliability, and tools for inspecting and controlling what AI agents do.
I'm especially interested in the parts around the model: preserving data through a workflow, understanding runtime failures, and making decisions traceable to code and policy. Most of my work is in Python, with JavaScript for MCP tooling and a particular interest in Apple Silicon.
- trustcard — tooling for inspecting MCP servers, signing capability manifests, detecting changes, and enforcing which tool calls are allowed. Available as
mcp-trustcard. - Apple Quality Recognition Engine — an apple-grading prototype that separates visual observations from a versioned grading policy. The current work centers on annotation, case review, and human-approved training data. It does not yet include a trained apple checkpoint or a validated commercial grading system.
These are merged contributions to projects maintained by their respective teams.
| Project | Contribution |
|---|---|
| Roboflow Inference | Reuse a run-scoped thread pool for workflow execution |
| Roboflow Inference | Preserve image metadata on empty Template Matching results |
| Roboflow Python SDK | Return upload results to callers |
| TensorFlow | Fix XLA DenseBincount negative-input validation for runtime tensors |
| Ultralytics | Exclude Sigmoid from TensorRT INT8 quantization |
I like understanding a system well enough to explain why it failed. I'm building that judgment through debugging, code review, and maintaining my own tools.
- Start with a reproducible failure and keep fixes focused.
- Add regression coverage and report exactly what was tested.
- Keep model outputs, policy decisions, and human approval separate where the application calls for it.
Older repositories include experiments and archived work. Their READMEs describe their status; the two projects above are my current focus.