AITRAF: Aggressive Inline Trick Recognition and Feedback

AITRAF is a monorepo for inline skating trick recognition and performance feedback.

Packages

• packages/aitraf-core: shared runtime processing and model-input helpers.
• packages/aitraf-train: Hydra-driven data ops, label ops, preparation, training, evaluation, metrics, tracking, configs, and scripts.
• packages/aitraf-api: FastAPI inference service for demo video listing, trick classification, and trick AQA predictions.

Workspace Setup

1. Start the dev container.
2. Copy .env.example to .env and fill in required AWS, MLflow, data path, storage path, API token, and registered model URI values.
3. Install workspace dependencies:

task install

Workspace Commands

Root commands are workspace-level only:

task install
task lint
task format

Use task lint and task format for workspace-wide checks from the repo root.

Train workflows are exposed through the train: task namespace. See packages/aitraf-train/README.md for commands and workflow documentation.

API workflows are exposed through the api: task namespace:

task api:run
task api:test

See packages/aitraf-api/README.md for runtime environment requirements and endpoint behavior.

Shared Workspace Assets

• data/: lightweight manifests and annotation-derived inputs.
• storage/: generated caches, model assets, run outputs, and larger local artifacts.
• notebooks/: analysis notebooks.

Package Docs

• aitraf-core
• aitraf-train
• aitraf-api