Staff Forward Deployed AI Engineer at Google in Seattle. I work on low-latency LLM inference and production AI agents.
Merged upstream
- ggml-org/llama.cpp #30779:
llama-servernow visits parent slots before their child slots, so requests that share a prompt prefix no longer need a third forward pass. Released inb12188. - AI-Hypercomputer/maxdiffusion #476: skips the CUDA-only
nvidia-cuda-cccldependency on macOS, where no wheels exist, so local installs no longer fail on Apple Silicon.
Writing
- Inside System 1 Models: how a transformer stops writing text and starts providing calibrated decisions.
- Thinking, Fast and Slow for Production Agents: pairing open-weights System 1 decision models with an LLM so every model plays to its strengths.




