MEL // FROM THE WORKSHOP
Work
Builds coming out of the workshop — mostly local-first AI, from MLX fine-tuning on Apple Silicon to security-gated agents. Still battle-testing; poke around the repos and grab the signal on the home page for what ships next.
GATED SRE AGENT
WIPAI SRE agent for k3s homelabs — observe k8s events, diagnose via LLM, act only through human-approved gates. The eval harness ships before the agent.
AI ASSISTANT
WIPSecurity-first, Jarvis-style assistant: permission-gated tool calls with single-use JWTs, MCP tool fingerprinting, pgvector memory with RLS, and fully local voice.
PHI-2 SERVING
WIPLocal REST API serving Phi-2 behind FastAPI, with an optional MLX LoRA adapter — native MLX on Apple Silicon or portable PyTorch CPU in Docker. Sync and streaming generation.
RAG EXPERIMENT
EXPERIMENTLocal retrieval-augmented-generation system using ChromaDB, sentence-transformers, and Phi-2 on Apple Silicon.
ML EXPERIMENT
EXPERIMENTLoRA fine-tuning of Phi-2 and StarCoder2 on Apple Silicon using MLX.
AI CAT COMPANION
EXPERIMENTGamified ML cat-mood classifier with a Streamlit feedback loop.