LoopCloser

The problem

An AI project starts with raw data and with the experts who know what it means. Getting that knowledge out of their heads usually happens in spreadsheets, with no structure, no assignment rules and no traceability.

LoopCloser turns curation and expert knowledge extraction into a first-class, governed process.

What it does

• Hierarchy. Organisations (managed by a superadmin) contain products, which contain use cases. Each use case declares its type: text classification, document classification or agent feedback.
• Configurable use cases. Allowed classes, assignment mode (one expert or N experts per item), whether experts may add a class on the fly, and generic feedback criteria: numeric (bounded float), categorical, boolean, text.
• Expert workspace. A personal queue per use case and a review screen adapted to the item type; submitting moves straight to the next item, so experts stay in flow.
• Import & storage. Items and documents are imported per use case. Each organisation can send its documents to its own S3-compatible bucket (AWS, MinIO, Scaleway, OVH…). The backend relays files, so buckets stay private, and secrets are referenced by environment-variable name and never stored in the database.

How it is built

FastAPI + SQLAlchemy + Alembic (SQLite by default, PostgreSQL in production), React + TypeScript + Vite + MUI + React Query on the front. The frontend calls the API on its own origin through nginx or Vite, so there is no CORS to configure. Persistent volumes and a fixed-name database volume make upgrades safe.

Why it matters

Everything built later, RAG databases and agents included, rests on this first step. LoopCloser makes it structured and auditable, so expert judgement becomes data the organisation keeps.