Payment AI Feature Store & Fraud-Risk Intelligence Platform

Subtitle: Evidence-grounded GenAI investigation and Risk/Compliance validation for payment fraud-risk ML.

This project is a local-runnable, AWS-patterned, production-style reference framework for a Payments AI/ML role. It demonstrates a payment fraud-risk ML lifecycle:

public/proxy fraud data + synthetic ISO 20022 pacs.008/pacs.002
→ canonical payment-event schema
→ PySpark-compatible ETL and Spark-SQL-style validation
→ point-in-time offline/online feature-store artifacts
→ vanilla fraud baselines
→ calibrated approve/review/block decision policy
→ reliability specialty upgrades
→ SageMaker-compatible pipeline and model registry artifacts
→ model card, validation memo, control matrix, release decision
→ optional CFPB-style evidence-grounded GenAI investigation

Why this project exists

The project is designed to show core skills for applied AI/ML in payments:

• payment-domain feature engineering
• fraud-risk modeling with class imbalance and review-capacity constraints
• PySpark/EMR-compatible batch feature processing
• offline/online feature-store fundamentals
• leakage-safe point-in-time training sets
• SageMaker-style model lifecycle and approval artifacts
• Risk/Compliance-ready validation documentation
• evidence-grounded GenAI investigation as a supporting layer, not the core decision engine

Claim boundaries

This project does not use proprietary bank data, JPMorgan Chase data, real customer data, or production payment data. It uses small generated/public-style samples and, when available locally, synthetic ISO 20022 payment data. AWS artifacts are local, AWS-patterned, EMR-compatible, and SageMaker-compatible unless explicitly run on AWS.

Quick start

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt
make run-demo
make test
make reports
make audit
# Optional full local verification from a clean generated-artifact state
make verify

The default demo runs without external downloads, Kaggle credentials, AWS credentials, or proprietary data.

Core reports

After make run-demo, review:

• reports/feature_store_validation_report.md
• reports/baseline_vs_specialty_ablation_report.md
• reports/validation_memo.md
• reports/kpi_policy_readout.md
• reports/release_decision.md
• reports/public_dataset_extension_report.md
• reports/transformer_finetuning_smoke_report.md
• reports/deployment_readiness.md
• reports/data_quality_and_backfill_report.md
• reports/model_monitoring_baseline_report.md
• reports/explainability_report.md
• reports/batch_scoring_contract_report.md

Repository structure

src/adapters/              dataset adapters and canonical schema mapping
src/spark_jobs/            PySpark-compatible ETL with pandas fallback
src/feature_store/         offline/online feature store, PIT joins, parity checks
src/modeling/              vanilla baselines, calibration, model evaluation
src/specialty/             residual correction, risk envelope pruning, optional PBHT
src/simulation/            KPI-based decision policy simulator
src/rag_investigation/     evidence retrieval, case summary, guardrails
src/sagemaker_pipeline/    SageMaker-compatible pipeline and model registry templates
reports/                   generated evidence artifacts

One-command outputs

make run-demo builds sample data, features, models, specialty comparisons, registry metadata, RAG samples, and validation reports.

Docker / Kubernetes readiness

This repository now includes Dockerfile, docker-compose.yml, and Kubernetes scaffolding under k8s/. These artifacts are intended to demonstrate deployment readiness only. They do not imply production deployment or cloud execution.

Public benchmark extension

Configure optional public benchmark paths in configs/dataset_sources.yaml to run against user-supplied PaySim, Bank Account Fraud Dataset Suite, or IEEE-CIS files. The default make run-demo remains external-download-free and uses generated public-style fraud samples when those files are absent.

Hardening additions in the loop-optimized version

• Champion/challenger operating-model selection using fraud-review utility, not PR-AUC alone.
• Drift monitoring report using temporal train/test PSI checks.
• Small-data stress tests for amount shock, score-compression drift, and envelope-pruning contribution.
• Local/GitHub runnable audit with required-file, claim-boundary, large-file, and secret-like-file checks.
• Data-quality baseline and deterministic backfill idempotence validation.
• Model-Monitor-style baseline, local explainability diagnostics, and offline batch scoring contract.
• GitHub Actions workflow scaffold under .github/workflows/local-ci.yml, plus pyproject.toml for Python project metadata and pytest configuration.

Recommended review order: PROJECT_MANIFEST.md, then the five core reports in reports/.