Machine Learning / Evaluated demonstration
Separate changing inputs from declining model performance.
Actual recording · Explanatory English captions · Original synthetic data
Recorded demonstration of a functional local application. A pilot defines data, access and deployment; this website serves the video and screenshots.
THE PROBLEM
THE WORKFLOW
An Extra Trees regressor learns from simulated sensors, with split conformal prediction bands. Switch between a stable process, shifted inputs and changed quality. The dashboard checks feature distributions and observed outcomes separately. A CSV following the demo schema can also be analysed.
RECORDED EVIDENCE
| Check | Cases |
|---|---|
| Documented checks; synthetic data | 30 of 30 |
30 heldout scenarios: ten separate seeds, three controlled conditions per seed. All 30 classifications match expectations. Prediction metrics use another heldout set; they do not establish industrial accuracy. Timings measure the analysis function on CPU with the model ready, excluding HTTP and training.
WHAT WE COULD BUILD WITH YOU
Existing model, features, availability of real outcomes, analysis window and acceptable thresholds.
Fictional process; it does not diagnose causes or guarantee interval coverage under drift. On 300 heldout rows, MAE was 1.10 and band coverage was 88.3%. Performance alerts require observed outcomes.
Scope a Machine Learning pilot ↗
Explore demo
Trials → uncertainty → next design
Recommend trials using Bayesian optimisation and resource limits.
10 of 10 checks · Original synthetic dataDocumented checks; synthetic data
Explore demo
Spectrum → components → unfamiliar signal
Separate spectral components and highlight signals the model cannot explain.
99 of 100 checks · Original synthetic dataDocumented checks; synthetic data