case study / active
VOCO
A fully offline Windows agent. No API keys, no cloud, no excuses.
impact
Fully offline — no API key required
stack
Problem
Every AI assistant assumes cloud APIs and stable internet. I wanted private, local-first automation for constrained or air-gapped environments.
Approach
VOCO runs a Qwen3 model locally via Ollama. A TF-IDF + Random Forest router classifies intent and dispatches to specialised tools (browser, file system, OS). A closed-loop ReAct executor retries on failure and logs confidence, step traces, and retry counts.
Outcome
A working demo of private local AI automation. Reproducible eval pipeline for measuring router improvements.
Key decisions
ML router first: fast classification means the LLM only runs when necessary.
ReAct loop with explicit success criteria prevents runaway execution.
Structured eval log schema enables empirical measurement of routing accuracy.