cardiag

cardiag {doctor,train,diagnose,triage,clean,inspect,ingest,scrape,serve,...} [options] [args]

cardiag is an end-to-end audio machine-learning pipeline for triaging mechanical faults from audio recordings (e.g. phone clips or scraped social media). It isolates the relevant mechanical sound span, embeds it with a frozen CLAP model, and applies small calibrated linear classifiers.The tool reports:- Whether something sounds wrong (fault vs normal)- Rough location in the car (6 zones)- Ranked shortlist of likely parts (12+ families)- Honest `UNCERTAIN` when the input does not support a confident callIt is explicitly a triage aid, not a safety-critical or definitive diagnostic. All numbers are honest (grouped-by-video cross-validation) and the pipeline is designed to be inspectable.

Install from the repository:

$ git clone https://github.com/adam-s/car-diagnosis
cd car-diagnosis
uv venv && source .venv/bin/activate
uv pip install -e ".[scrape,web,dev,viz]"

Run `cardiag --help` or `cardiag <subcommand> --help` for details.Common workflow for a new clip:1. `cardiag doctor`2. `cardiag diagnose my-recording.wav --json`3. `cardiag inspect my-recording.wav -o report.html` (visual + audio spans)Training your own data:- `cardiag ingest ./myclips --kind fault --cause wheelbearing`- `cardiag scrape youtube|tiktok ...`- `cardiag train`The `serve` subcommand starts a local web UI for drag-and-drop or URL-based analysis. Add `--json` to inference commands for machine-readable output.

- Performance is modest on crude phone/social audio (literature ceiling for the problem). It reaches ~0.79 AUROC fault/normal and top-3 part accuracy 45-65 % in honest evaluation.- Models are joblib artifacts; only load ones you trust.- Not a replacement for professional diagnosis.

Source code · Documentation