audit()Red threadID columns, duplicates, label noise, and single-feature target leakage, caught before you train.
The honesty room
Numbers, with their caveats
The site refuses to glamorize bad ML, and it refuses to glamorize its own benchmarks.
BreezeML vs PyCaret vs LazyPredict
Measured 2026-07-05, Windows 11, Python 3.11, same machine and venv. Wine dataset.
| Library | Install | Import | Leaderboard | Accuracy | LOC | Lock-in |
|---|---|---|---|---|---|---|
| BreezeML | 2m 36s / 274 MB | 3.1s | 9.0s | 1.000 | 3 | no |
| PyCaret | 6m 21s / 952 MB | 6.9s | 19.3s | 0.984 | 5 | yes |
| LazyPredict | (shared venv) | 7.2s | 3.7s | 1.000 | 8 | yes |
- LazyPredict's leaderboard is genuinely the fastest, but it returns no usable pipeline: no preprocessing, no persistence, no deployment.
- Wine is a small, clean dataset. On large, messy problems a tuned PyCaret stack or AutoGluon may find more accurate ensembles.
- BreezeML's claim is not 'most accurate AutoML'. It is the fastest path from DataFrame to a sound, explained, deployable model, with 4 dependencies and zero lock-in.
fairness.report()Balance bridgePer-group metrics, demographic parity, and a four-fifths rule verdict.
drift.check()Weather changePSI drift, unseen categories, and range violations, with a live /drift endpoint in every deployed API.
significanceTwo bellsMcNemar and paired-CV t-tests: is that 2% gain real, or noise?
conformalMeasured lanternDistribution-free prediction intervals with a guaranteed coverage level. Honest uncertainty on every prediction.
card()Model scrollAn honest model card: data profile, metrics, decisions, and auto-detected caveats.
Zero lock-in, proven
export() writes a standalone scikit-learn script that reproduces your exact pipeline, with no breezeml import in it. Run it, delete BreezeML, and the script still works.
python benchmarks/run_benchmarks.py