Turn any CSV into a production-ready model and know exactly how and why, every step of the way.

Most dev teams sit on data worth predicting from and never act on it. Not because the idea isn't there, but because building a reliable ML pipeline is a project in itself: weeks of preprocessing decisions, model selection guesswork, hyperparameter tuning, and evaluation work, all before you've written a single line of serving code.

FlowML removes all of that.

Upload a dataset. Pick a target column. A team of AI agents handles the rest (analysis, feature engineering, model selection, training, evaluation, and deployment) while logging every decision it makes so you can inspect, trust, and ship the result.

And when the pipeline finishes, you're not done interacting with it. A conversational chatbot agent sits on top of the entire system, letting you modify, rerun, and interrogate the pipeline through natural language. Ask why a feature was selected. Undo a change. Compare two models. It plans the revision, executes the affected stages, and tracks the full history, so every experiment is reproducible and nothing is lost.

No ML background required. No black box you can't interrogate. No hand-off to a specialist.

1. Upload a dataset: drag and drop a CSV into the browser
2. Pick a target column: the column you want to predict
3. Run the pipeline: eight agents execute in sequence, visualised in real time
4. Inspect every decision: click any stage to see what the agent did and why
5. Refine conversationally: tell the chatbot what you want to change; it plans the revision, reruns only the affected stages, and stores the diff
6. Ship the result: download a deployment package with a working REST API, Docker config, and a plain-English model report ready for stakeholders

A floating chat interface gives you full control over the pipeline through natural language. This isn't just a Q&A wrapper-- it's an orchestration agent that can modify, rerun, and audit the pipeline safely.

A successful pipeline produces a ready-to-ship zip containing:

deployment/
  app.py              ← FastAPI serving endpoint, schema-validated
  requirements.txt    ← pinned to training versions
  Dockerfile
  docker-compose.yml
  schema.json         ← column names, types, expected ranges
  README.md           ← What the model does, caveats, input/output spec
  model.pkl           ← trained model artifact 
  report.html         ← styled pipeline report with metrics and explanations

docker-compose up and your model is serving. The generated README.md transfers knowledge automatically, removing human dependency.

• FastAPI
• pandas / NumPy
• scikit-learn
• XGBoost / LightGBM
• Optuna (HPO)
• SHAP

• React 18 + TypeScript + Vite
• Tailwind + Radix UI
• Recharts
• Framer Motion

agents/        Specialized AutoML agents (analysis, preprocessing, training, etc.)
api/           FastAPI endpoints and pipeline stage execution
core/          Orchestration and ML utilities (comparison, HPO, ensembling)
frontend/      React UI and dashboards
utils/         Shared helpers (logging, OpenRouter client, insights)
tests/         Pytest test suite
outputs/       Saved artifacts (models, metadata, reports, deployment zips)
uploads/       Uploaded datasets (deleted after pipeline completes)

pip install -r requirements.txt

Copy .env.template to .env and set OPENROUTER_API_KEY:

uvicorn api.main:app --reload --port 8000

cd frontend
npm install
npm run dev

The frontend targets http://127.0.0.1:8000 by default.

pytest
pytest tests/test_agents.py -v   # specific file

cd frontend
npm test
npx tsc --noEmit   # type check