Credit‑Scoring AI‑Act Demo

Demo project to showcase EU AI Act compliance artifacts with ZenML.

Project Structure

credit_scoring_ai_act/
├── data/
│   ├── raw/ # public synthetic loan data (CSV)
│   └── snapshots/ # auto‑versioned by data_versioning step
├── pipelines/
│   ├── credit_scoring.py # "happy path" train‑→‑deploy
│   ├── monitor.py # scheduled drift checker
│   └── retrain.py # triggered by incident label
├── steps/
│   ├── data_loader.py # load CSV → log SHA‑256, WhyLogs profile
│   ├── data_preprocessor.py # basic feature eng, outputs train/test splits
│   ├── data_splitter.py # split dataset into train/test
│   ├── generate_compliance_metadata.py # generate compliance metadata
│   ├── train.py # XGBoost / sklearn model
│   ├── evaluate.py # std. metrics + Fairlearn/Aequitas scan
│   ├── approve.py # human‑in‑loop gate (approve_deployment step)
│   ├── post_market_monitoring.py # post‑market monitoring
│   ├── post_run_annex.py # generate Annex IV documentation
│   ├── risk_assessment.py # risk assessment
│   └── deploy.py # push to Modal / local FastAPI
│
├── compliance/
│   ├── templates/
│   │   ├── annex_iv_template.j2 # annex iv template
│   │   └── fria_template.md # rights‑impact narrative
│   ├── records/ # automated compliance records
│   ├── manual_fills/ # manual compliance inputs
│   ├── monitoring/ # monitoring records
│   ├── deployment_records/ # deployment history
│   └── approval_records/ # approval history
│
├── reports/ # auto‑generated Docs, PDFs, JSON logs
│   ├── annex_iv_<run>.pdf
│   ├── model_card_<run>.json
│   └── incident_log.json
├── utils/ # shared utilities
├── configs/ # configuration files
└── README.md

Pipeline graph

1. ingest -> preprocess -> train -> evaluate -> approve_deploy -> deploy
2. Scheduled: monitor (daily) → report_incident on drift.
   Incident closed? → GitHub label retrain triggers retrain pipeline.

Where each compliance artifact is produced

Step / Hook	AI‑Act Article(s)	Output artefact
ingest	10, 12	dataset_info metadata, SHA‑256
preprocess	10, 12	preprocessing_info_*.json, data quality logs
evaluate	15	fairness_metrics, metrics.json
approve_deploy	14	approval.log (approver, rationale)
post_run_annex.py	11, 12	annex_iv_<run>.pdf
monitor.py	17	drift scores in reports/drift_<date>.json
incident_webhook.py	18	incident_log.json, external ticket
risk_assessment.py	9	risk scores, risk register updates

Compliance Directory Structure

Directory	Purpose	Auto/Manual
records/	Automated compliance records from pipeline runs	Auto
manual_fills/	Manual compliance inputs and preprocessing info	Manual
monitoring/	Post-market monitoring records and drift detection logs	Auto
deployment_records/	Model deployment history and model cards	Auto
approval_records/	Human approval records and rationales	Manual
templates/	Jinja templates for documentation generation	Manual

"Definition of Done"

zenml up && zenml run credit_scoring
# -> reports/annex_iv_<run>.pdf exists
# -> fairness_metrics logged
# -> approval gate records reviewer
# -> drift monitor cronjob scheduled
# -> compliance/records/ contains run artifacts
# -> compliance/manual_fills/ contains required inputs