The problem
I kept needing labeled data and every existing annotation tool was either (a) a SaaS with per-seat pricing that academic budgets can't handle, or (b) a local tool that required a PhD in configuration to run.
I wanted something you could clone, run with python app.py, and be annotating in under 2 minutes.
What I did
Built a minimal annotation tool focused on classification and NER tasks. FastAPI backend with SQLite (no Postgres setup required), React frontend with keyboard shortcuts for everything, and a simple CSV/JSONL export.
Key design decision: no user authentication. If you're using this, you're a researcher running it locally. Auth adds complexity without adding value in that context.
Added multi-annotator support later — you can assign tasks and compute inter-annotator agreement (Cohen's kappa) directly in the UI.
What I learned
SQLite is underrated for tools like this. It's a single file, you can email your database, and for anything under 10GB it's fast enough that you'll never notice the difference from Postgres.
The keyboard shortcut investment paid off enormously. Once annotators memorize 5 shortcuts, their throughput roughly doubles. Time spent on that UX detail had the highest ROI of anything in the project.
Two labs ended up using it for unrelated projects, which was validating — it meant the scope was right.