Product · 2026
ShopFloor Copilot
A manufacturing operations copilot that plans its own investigation in plain English, writes guarded SQL against a 16-table plant database, and never takes an action without human approval. Built on Gemini 2.5 through Vertex AI, with a full audit trail and a hard monthly budget guard.
The live product is private by design.
The running copilot calls a paid LLM on every turn, so it sits behind an access-code gate that fails closed — unauthenticated requests are rejected before any model or database call, at zero cost. The demo above is the real product, recorded end to end.
What it does
Ask a plant-floor question in plain English — "what's the yield difference between Line 3 and Line 5 last week, and why?" The copilot drafts an investigation plan and pauses for a human to run, edit, or cancel it. Once approved, a tool loop resolves the entities, writes read-only SQL, and composes an answer that cites the exact rows behind every number. Every statement, decision, and action is recorded.
How it's built
- Planner + approval gate. A Gemini 2.5 planner (temperature 0, for reproducibility) proposes the investigation; nothing runs until a human approves.
- Four-layer SQL guardrail. Every generated query passes an AST table whitelist, read-only execution, a statement timeout, and an injected row limit before it touches data.
- Grounded answers. The synthesizer may only cite values the tools actually returned — the UI shows the evidence rows beside the conclusion.
- Observability & durability. Full Langfuse traces per turn, an eval harness in CI, SQLite replicated to object storage, and keyless deploys to Cloud Run.