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 and coordinate contracts through a workflow, isolating runtime failures across hardware backends, and making automated decisions traceable to code and policy. Most of my work is in Python, with JavaScript for MCP tooling and a focus on Apple Silicon.
- trustcard — tooling for inspecting MCP servers, signing capability manifests, detecting capability drift, and enforcing which tool calls are allowed. Published as
mcp-trustcard.
Merged pull requests in projects maintained by their respective teams:
| Project | Contribution |
|---|---|
| Roboflow Inference | Reuse a run-scoped thread pool for workflow execution |
| Roboflow Inference | Preserve image dimensions on empty VLM detections |
| Roboflow Inference | Clear every touched namespace in workflow cache blocks |
| Roboflow Python SDK | Return single-upload results from Project.upload() |
| TensorFlow | Fix XLA DenseBincount negative-input validation for runtime tensors |
| Ultralytics | Exclude Sigmoid from TensorRT INT8 quantization |
| Hugging Face PEFT | Document hybrid Mamba-Transformer target_modules pitfalls |
All merged upstream pull requests
I focus on understanding a system well enough to reproduce and explain why it failed—building that judgment through root-cause debugging, differential testing, and maintaining my own tools.
- Start with a minimal reproducible failure and keep fixes narrow.
- Add regression tests that fail before the patch and report exact verification boundaries.
- Keep model observations, deterministic policy, and human authorization separate where reliability matters.
Older repositories hold prototypes, archived experiments, and retired research (including apple-quality-recognition-engine, closed out with a documented engineering postmortem and partial dataset). Each repository's README states its current status.



