Attune is a mixed-initiative system for scoring text records at scale with LLMs. Instead of scoring records in isolation with a prompt, Attune compares records pairwise to build a holistic understanding of the data, resolves those comparisons into consistent scores, and derives the scoring criteria and rules bottom-up — then lets you inspect and deterministically refine that scoring logic through direct manipulation and natural-language feedback.
Live deployment: attune-alpha.vercel.app
Prerequisites: Python ≥ 3.12, Node.js ≥ 20.9.
cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
uvicorn app:app --reload --port 8000cd frontend
npm install
npm run devIn the frontend, open Edit › Edit API Keys, set a LiteLLM model string (e.g. openai/gpt-4.1, azure/<deployment>, anthropic/claude-sonnet-4-5), and paste the matching provider key (OpenAI, Anthropic, Azure, Gemini, OpenRouter, or any custom variable LiteLLM understands). Keys are not stored and need to be re-entered for every browser session.
If you are running locally and wish to avoid re-entering your key, you can set them in backend/.env (see backend/.env.example).
To get started, pick a sample dataset from the dropdown or Upload your own dataset. Use the Task Configuration pane to describe the scoring task, choose the field(s) to score, mention the score range, and hit Run Initial Scoring.
Uploading Data: Upload a JSON array of flat records.