Building intelligent systems that run on your hardware, with your data, under your control.
I architect local-first AI orchestration systems — multi-agent frameworks that run entirely on consumer hardware (Apple Silicon) with zero cloud dependency. My work spans the full stack from RAM-aware model scheduling to terminal-native REPLs, from vector databases to no-code workflow platforms.
An evolving exploration into hardware-aware AI orchestration. Six iterations, each pushing the boundary of what's possible locally.
| Version | Focus | Key Innovation |
|---|---|---|
| v6 — Mesh | Visual Mesh + Bot Gateway | No-code AI orchestration via Docker, Ollama, Dify |
| v5 | Lazy Conductor | Trio structured concurrency + Ponytail minimal-code philosophy |
| v4 | Multi-Backend Orchestrator | MLX · Ollama · Cloud · Hybrid parallel execution + REPL |
| v3 | Portable Hot-Swap | Single model slot with adaptive planning loop |
| v2 | Sequential Orchestration | 4GB coordinator + 20GB hot-swap slot on Apple Silicon |
| v1 | MLX Multi-Agent | Brain LLM plans → delegates → reviews until production quality |
Scans any GitHub repo and scores extractable components worth shipping as standalone packages.
Supported: 13 languages · 6 scoring dimensions · 286 tests · battle-tested on Twitter, Flask, React, Kubernetes, CPython (58K+ files, zero crashes). Includes extraction scaffolding, security hardening (symlink protection, secret redaction, git filter RCE mitigation).
Repository Composition Engine — discover, evaluate, and generate blueprints to combine existing open-source tools.
Evolved from GitHub Repo Finder. Task-to-capability mapping, multi-dimensional health scoring, automated recipe builder, AI-agent token optimization.
Production-grade GitHub Repo Finder for LLMs — with MCP and skill file support.
Three specialized assistants sharing a unified core framework + ChromaDB vector store:
| Assistant | Role | Stack |
|---|---|---|
| NAMI | System Control & Monitoring | Telegram bot · Remote execution |
| RUSH | Audio Transcription & Summarization | whisper.cpp (Metal) · Ollama |
| VEX | Video Scene Detection & Clipping | PySceneDetect · ffmpeg |
Includes Flask dashboard, launchd auto-start, and shared knowledge graph.
Systems Architecture:
RAM-aware scheduling · model hot-swapping · circuit breaker patterns · structured concurrency (Trio) · pure Python state machines · event-driven pipelines
AI/ML Stack:
MLX · Ollama · ChromaDB · Weaviate · Whisper.cpp · sentence-transformers · RAG pipelines
Infrastructure:
Docker · launchd · Flask · python-telegram-bot · httpx · Dify · n8n
"The best AI is the one that runs on your hardware, with your data, under your control."
Local-first AI isn't a compromise — it's a design choice that unlocks privacy, latency, and cost advantages cloud AI can't match. Every iteration of my orchestration engines has been a step toward making local AI systems as capable and easy to use as their cloud counterparts, using a fraction of the resources.
Built with obsession for clean architecture and zero-bloat systems.