07 / Independent ML systems project · 2026
Environmental audio clustering as a reproducible system
One canonical feature contract from waveform generation to served cluster assignment.
The original README-only repository is now an executable system. Training and serving share an explicit feature schema, while synthetic signals provide a licensed, deterministic public demo path.
Evidence register
170
features
Canonical train/serve schema
4/4
tests
Local and CI suite
ARI
external metric
Computed only when labels exist
01 / Problem
Notebook-only clustering cannot prove that feature extraction, training and inference agree on the same representation.
02 / Approach
A shared extractor validates audio, builds 170 features and feeds both the Kedro training pipeline and bounded inference API.
03 / Outcome
The repository now runs end to end with generated signals, packaging, tests, CI and container documentation.
System trace
How the evidence is produced.
Rebuilt the repository into a reproducible package, test suite, CI workflow, container and bounded API.
- 01Waveform → validated audio
- 02Shared extractor → 170 features
- 03Pipeline → clusters and artifact
- 04API → bounded assignment
Validation scope
Four tests cover feature shape, deterministic generation and core pipeline contracts. Synthetic labels allow a limited ARI sanity check.
Known limitation
Synthetic separability does not establish usefulness on a real acoustic corpus; stability and representation quality still require external evaluation.
What is inspectable
- Train and serve import one schema.
- The public demo needs no opaque dataset.
- ARI is never fabricated for unlabeled data.
Next proof to add
- 01Evaluate on a governed public corpus.
- 02Measure stability across seeds and perturbations.
- 03Add latency and drift monitoring.
Main stack