I work at IBM Research. I love building agents, loops, and agents that loop. Sometimes they even converge.
I also enjoy algorithms, LLM inference, GPU kernels, retrieval, and bugs that turn into interesting benchmarks.
- llama.cpp - Started playing with local models on my own hardware, got curious about making them faster, and somehow ended up with a few merged PRs. Mostly Metal kernels, MoE optimizations, and the occasional bug fix.
- Spotlights - local evidence retrieval and multi-CLI research workflows.
- oh-my-codex - a Stop-hook reliability fix.
-
🧙 NanoSage - I wondered if deep research really needed heavy compute, or if an agent could explore questions, follow evidence, and write useful reports using a 2B Gemma model running privately on my laptop. Built NanoSage to find out. Example report.
-
🐍 Codex Python SDK prototype - I wanted to puppet Codex from inside an agent sandbox and see if a coding agent could run deep research workflows, not just write code. Ended up building an unofficial Python interface with streaming, thread management, and structured output. Upstream discussion.
-
⚡ tfrq - a small Python helper for running independent functions concurrently.
-
Qiskit - transpiler bug reports (#17314, #17315, #17316) and compiler experiments in my fork (#2, #3).
-
vLLM - Some speculative decoding and sampler memory experiments. Experimental and historical now, lol.
I'd love to work with you. Bring an interesting bug, a weird idea, or a benchmark worth chasing.
Humans are welcome too, obviously. 😄
LinkedIn | IBM Granite Vision 3.3 (2B) Embedding - model card