AI infrastructure engineer. I make model inference faster on phones, laptops, and cloud GPUs.
Until November 2026 I was at RunAnywhere (YC W26). I built hybrid routing and the agent gateway in Wally, cut inference latency on the backend, and worked on the SDK through llama.cpp and MLX.
ToolNeuron. Offline Android AI. Chat, images, and speech stay on the device. 16K+ installs.
Ai-Systems-New. The C++ SDK behind ToolNeuron. llama.cpp for chat, QNN and MNN for images, ONNX Runtime for speech.
llama.cpp-android. llama.cpp for Android, with ARM kernels and big.LITTLE scheduling. About 50 tokens/s on one Cortex-X3 core.
ForgeAI. Desktop app to load, inspect, and merge model files. Rust and Tauri.
I also wrote Hexagon DSP kernels that skip the QNN SDK. The benchmark reported 8 TFLOPS. The matrix unit was fused off. Writeup.
C++ Rust Kotlin llama.cpp GGML MLX Android Hexagon





