Situation
Strong held-out accuracy does not reveal when a classifier stops working in production. Teams need evidence of where a model remains reliable, not just a single number.
Data Science
FaultForge
A model-evaluation workbench that measures how classification pipelines behave under missingness, noise, distribution shift, and other real-world stresses.
View source repository↗
Why it matters
Replaces a single accuracy score with reproducible evidence showing where a model remains reliable and where it does not.
My role
Data scientist, model-evaluation designer, and solution developer
- 01Defined reproducible stress plans, metric guards, deployment ranges, adequacy rules, and confirmed breaking-point logic.
- 02Implemented repeated-seed evaluation, paired bootstrap evidence, calibration, prediction margins, flip diagnostics, PSI screening, and comparable version ranking.
- 03Designed evidence outputs that distinguish pass, performance failure, execution failure, flakiness, and inconclusive results instead of forcing false certainty.
Task
Build a model-evaluation workbench that deliberately stresses a classifier with missingness, numerical noise, distribution shift, categorical drift, unseen categories, and prevalence change, then turns the results into reproducible, honest pass or fail evidence instead of false certainty.
Action
- Every severity level runs across multiple deterministic seeds, reporting mean, standard deviation, minimum, and maximum. Conservative guards compare mean minus standard deviation against user-defined requirements.
- A failure only becomes a confirmed breaking point once all later severities also fail. Isolated drops are marked flaky, and the threshold is reported as an interval rather than a false-precise number.
- The baseline covers accuracy, balanced accuracy, recall, precision, F1, ROC-AUC, PR-AUC, and Brier score, backed by paired class-stratified bootstrap evidence so unsupported drops are never promoted as confirmed failures.
- Calibration, probability margins, per-row prediction flips, and PSI feature-drift rankings diagnose how degradation actually appears, while immutable model and dataset fingerprints keep every decision traceable to its exact inputs.
Result
Deployment ranges and adequacy rules (fewer than 30 minority-class examples is inconclusive, fewer than 10 is rejected) keep promotion decisions honest. A single accuracy score is replaced with reproducible evidence of exactly where a model remains reliable and where it does not.
Technical implementation
How the solution was built.
Each layer connects an implementation choice to the decision or workflow it supports.
06 layers| Layer | Implementation | Operational purpose |
|---|---|---|
| Experiment definition | Versioned model artifacts, dataset fingerprints, target semantics, stress plans and metric guards | Make every evaluation reproducible and tied to the exact model and data used |
| Stress generation | Missingness, numerical noise, distribution shift, categorical drift, unseen categories and prevalence change | Measure behaviour under realistic data-quality and population changes |
| Repeated evaluation | Deterministic multi-seed trials with mean, standard deviation, minimum and maximum summaries | Separate stable degradation from one-off variation |
| Statistical evidence | Paired class-stratified bootstrap intervals and conservative guard values | Avoid declaring failure when the observed difference is not adequately supported |
| Diagnostics | Calibration, confusion matrices, probability margins, prediction flips, PSI and compound-stress ablations | Explain how and where model behaviour changes |
| Decision output | Pass, failure, flaky, inconclusive and execution-failure states with JSON and HTML evidence | Translate evaluation results into reviewable promotion decisions |
Product walkthrough
Screens connected to decisions.
robustness confusion calibration
Combines confusion-matrix, calibration, and robustness curves into a single evaluation view.
stress severity results
Shows how a metric degrades across increasing stress severity levels, with confirmed breaking points.
model registry
Organizes versioned model assets and keeps lineage and reliability evidence attached to each model.
model comparison
Compares reference and challenger versions using evidence relevant to promotion decisions.
new experiment
Turns a robustness evaluation into a guided, reproducible configuration workflow.
model detail
Documents a single model version's lineage, configuration, and evaluation history.
stress matrix
Maps pass, fail, and flaky outcomes across every stress type and severity level tested.
seaborn line
Plots metric trends across severity levels to visualize where performance actually breaks down.