Offline-first guardrail that sits between an AI agent and its MCP tools. Every proposed call is normalised, policy-checked, audited, and answered ALLOW / REVIEW / BLOCK before any side effect.

Result, on 220 held-out MSB cases (qwen2.5:7b): harmful actions executed fell from 58.5% → 0.0%. The repo is upfront that a hold-everything baseline also hits 0.0%, so it reports what IntentShield does differently (no harmful call reaches the review queue) and what it costs (9–14 pts task completion).

Python MCP Policy engine uv

Bearing-fault detection from vibration + audio. A small local detector flags abnormal windows; a bounded LangGraph agent (≤2 retrievals, ≤2 LLM calls) drafts a cited inspection report that only an authenticated human can approve.

Result: 0.955 ± 0.079 balanced accuracy (fusion Isolation Forest, leave-one-bearing-out over 15 held-out bearings), with the weaker cross-rig transfer reported too.

Python LangGraph FastAPI scikit-learn

Comment-to-DM automation for Instagram and Telegram with a native React Native operator app. Per-account rate limiting keeps it under Meta's documented 750 private replies/hour cap. CI runs typecheck, lint, tests and build.

Next.js 16 React Native Prisma BullMQ Redis PostgreSQL

QLoRA-style fine-tune of Qwen2.5-1.5B (4-bit) trained locally on an Apple M5 with MLX; base vs fine-tuned compared on 200 held-out examples under identical decoding.

MLX LoRA Python