I build local AI systems and tools for working with AI agents—from model-specific inference runtimes to speech interfaces, multi-agent apps, and desktop robots.
- llm-in-c — Model-specific runtimes and hardware optimizations for text, speech, and video generation. Native C and Metal implementations, with benchmarks organized by model and hardware.
- whisper-turbo.c — Whisper large-v3-turbo transcription in C, with INT8 weights and AVX2/AVX-512 CPU acceleration. A command-line runtime focused on low-memory speech recognition.
- Voice Feedback for Claude Code + Codex — Local voice-cloned feedback on Apple Silicon using MOSS-TTS-Nano and MLX. Shared playback queues keep concurrent agents from speaking over one another.
- Agentrooms — Native macOS and iOS app for coordinating coding agents across local and remote machines, with threaded conversations and direct agent mentions.
- Claude Remote Proxy — Run Claude Code remotely while keeping the interactive terminal local, with workspace and session synchronization over SSH.
- Yuki — An expressive desktop robot with on-device face detection, gesture-triggered conversations, and coordinated animation, movement, and light.
Tools that connect AI assistants to applications and services:
- Remote macOS control — screenshots, keyboard, and mouse control for remote Macs.
- HubSpot — CRM access with vector search and caching.
- OpenMemory — persistent conversation memory.
- OpenAPI tools — API discovery and execution through semantic search.