← All projects
Orange Business · 2026 — present · live

Vigie

An agent that reads a team's documents — meeting minutes, specifications, tenders —, answers while citing its sources, then acts in the workspace: project created, task added, meeting scheduled. Live since August 31, 2026, by personal invitation: around ten users to date.

Impact
0 actions taken without human validation, by design
Role
AI Product Engineer
Stack
Agent LLM, React, Node, PostgreSQL, OpenRouter, Security
Keywords
Agent, Documents, Private AI Factory

The problem

An assistant that only answers leaves all the follow-through to the user: rereading the minutes, extracting the decisions, creating the tasks, scheduling the meeting. An assistant that acts on its own, on the other hand, becomes impossible to hand over to a team — nobody signs off on a tool that writes into projects without anyone knowing what or when. Vigie is built to do both: it acts, but no action goes out without explicit approval.

The approach

A 138-line agent loop that chains reasoning, tools and response, six turns at most. Nine tools declared in a registry where each one states whether it reads or modifies — it is this declaration, not a rule written somewhere else, that triggers the validation request. The model has no access to the application: it requests a tool by naming its arguments, the browser executes. A refused action is sent back to it as is, and it takes that into account on the next turn.

All data goes through a single access point, which makes the database replaceable, and a single gateway knows the model provider: switching it only touches one file.

Security, tested before being announced

Access by personal invitation, passwords hashed with scrypt and a per-account salt, rate-limited login. An uploaded document is data to read, never an order to follow: untrusted content is fenced off and any injection attempt is flagged rather than silently executed. Database hosted in Ireland, encrypted at rest; providers that retain requests are refused on every call.

Every protection has been put to the test — access attempted without an account, passwords brute-forced one after another, booby-trapped documents uploaded to see what happens.

Cost, enforced by the server

Every call is counted in real tokens and real cost, with a cap of $10 per person per month: alert at 80%, block at 100%, reset on the 1st. With the chosen model, those $10 represent roughly fifty thousand questions — the cap protects against runaway usage, it does not ration it. The control is enforced server-side, so it cannot be bypassed from the browser.

Where it stands

Around ten people have access today, each invited by name.

  • In place: the agent and its tools, traceability, the foundation and the database, budgets and quotas.
  • In progress: fine-grained access control and user-selectable model choice.
  • To do: connectors to existing tools — email and calendar — which will go through MCP.

Two building blocks remain in-house, the agent loop and the orchestration, because no off-the-shelf assembly would do better at this scale; authentication, for its part, is set to move to Supabase Auth.