I build LLM agents and the evaluation around them. I'm a Senior AI Engineer at HP's AI Center of Excellence.
- Now: at HP I led the design and implementation of an internal typed agent library (Pydantic AI and Pydantic Graph), and led the implementation of Nio, a customer-facing assistant, from prototype to general availability.
- Before: I was the first hire at Methinks, a stroke-imaging startup, and stayed eight years: detection of large-vessel occlusion and hemorrhage from non-contrast CT, CE marked, FDA cybersecurity certified, deployed in hospitals, validated in Stroke. Two Kaggle RSNA gold medals with the team.
- At home: a DGX Spark serving local models behind llama-swap, where I test inference and quantization.
- inspect_ai#5799: name scores with a dict of scorers in Inspect's
score()(in review). - vllm#59770: reproduced an NVFP4 decode regression on GB10 and traced it to a FlashInfer cuDNN gate.
- agent-harness-path: a self-contained course on building, evaluating and governing LLM agents.
- rayuela: a computational reading of Cortázar's Rayuela with embeddings, LLMs and permutation tests.
- docker-neural-memory: neural memory that learns at test time (Titans / TTT), in a container.
Python · PyTorch · Hugging Face · LLM agents (Pydantic AI, MCP) · evals and tracing (Langfuse, MLflow) · vLLM · llama.cpp · Docker / Kubernetes · AWS