Mobile AI, ternary quantization, C++ and Rust.
I build tools for running large language models on phones.
- PEHT folded Hadamard ternary 27B models on Android, 13x faster than the vendor PQ2_0 path at the same quality
- phone-cluster one local LLM across two Android phones over USB with ggml-rpc
- upstream work in llama.cpp Vulkan device fault handling and ggml-rpc reliability
Focused on ternary weights, phone memory bandwidth, and NPU/CPU scheduling.

