Banking Fraud Demo¶

The banking-fraud demo is AKKO's reference end-to-end use case. It exercises every layer of the platform — ingestion, lakehouse, compute, ML, governance, BI, AI — in under 10 minutes on a laptop-sized k3d cluster, 100 % offline.

What It Proves¶

Airflow → Iceberg (Polaris) → Trino federation → dbt marts
MLflow experiment tracking + model registry backed by MinIO
OPA column masking on PII (card_pan, customer_name) per Keycloak role
ADEN natural-language question answering on the curated layer
Trino AI functions (akko_ai_anomaly, akko_ai_sentiment) inside a single SQL
Superset dashboard auto-provisioned by the akko-init job
Tempo end-to-end tracing of the Airflow → Trino → ADEN span

Architecture¶

flowchart LR
    GEN[generate_transactions.py<br/>Faker + PostgreSQL seed]
    AF[Airflow DAG<br/>banking_fraud_demo]
    SPK[Spark Connect<br/>JDBC → Iceberg MERGE]
    POL[Polaris<br/>Iceberg REST catalog]
    MINIO[(MinIO<br/>akko-warehouse)]
    TRI[Trino 480<br/>+ ai_* plugin]
    DBT[dbt Core<br/>marts/fct_transactions]
    ML[MLflow<br/>Isolation Forest]
    OPA[OPA<br/>PII mask per role]
    SUP[Superset dashboard<br/>AKKO Banking Fraud]
    ADEN[ADEN<br/>NL → SQL → Streamlit]

    GEN --> AF --> SPK --> POL --> MINIO
    POL --> TRI --> DBT --> POL
    TRI --> ML
    TRI --> SUP
    TRI --> ADEN
    OPA -.policy.-> TRI

Run It¶

# 1. Deploy AKKO on k3d
bash helm/scripts/generate-domain-values.sh akko.local
bash helm/scripts/generate-dev-secrets.sh
bash helm/scripts/deploy.sh

# 2. Trigger the demo DAG
kubectl exec -n akko deploy/akko-airflow-scheduler -- \
  airflow dags trigger banking_fraud_demo

# 3. Watch the run
kubectl exec -n akko deploy/akko-airflow-scheduler -- \
  airflow dags list-runs -d banking_fraud_demo

# 4. Open Superset
open https://bi.akko.local/superset/dashboard/akko-banking-fraud/

Pipeline Steps¶

Seed — generate_transactions.py inserts 100 000 transactions into postgresql.akko.raw_transactions using Faker.
Replicate to Iceberg — Spark Connect reads from PostgreSQL and does MERGE INTO iceberg.banking.transactions partitioned by region.
dbt models — dbt run --models marts.fct_transactions marts.dim_customers rebuilds curated layers.
Score anomalies — an Isolation Forest is trained in a JupyterHub notebook, logged to MLflow with artifacts on MinIO, and registered as banking-fraud/production.
Enrich with AI — a single SQL runs akko_ai_anomaly(amount, ...) and akko_ai_sentiment(memo) to enrich each row.
Refresh dashboard — Superset dataset is backed by the Trino view; cache invalidates on DAG success.
Ask ADEN — "Which regions saw a fraud rate above 5 % last week?"

RBAC Matrix on `iceberg.banking.transactions`¶

Column	`akko-admin`	`akko-engineer`	`akko-analyst`	`akko-user`	`akko-viewer`
`customer_id`	clear	clear	clear	clear	clear
`customer_name`	clear	clear	clear	`*MASKED*`	`*MASKED*`
`card_pan`	clear	clear	`*MASKED*`	`*MASKED*`	`*MASKED*`
`amount`	clear	clear	clear	clear	clear
`region`	clear	clear	clear	clear	row filter (EU/APAC)
`fraud_score`	clear	clear	clear	clear	clear

Row filter: akko-viewer only sees EU/APAC regions with status = 'active'.

Files in the Repo¶

File	Role
`airflow/dags/akko_banking_fraud_demo.py`	Airflow DAG orchestrating the full pipeline
`airflow/scripts/generate_transactions.py`	Faker-powered synthetic transactions
`notebooks/akko-banking-demo.ipynb`	Isolation Forest training, MLflow logging
`dbt/models/marts/fct_transactions.sql`	Fact table over the Iceberg base
`superset/assets/bootstrap_dashboard.py`	Auto-provisioning of dataset + charts + dashboard
`keycloak/realm-akko.json`	Keycloak roles + clients used by the demo

Troubleshooting (Top 5)¶

Symptom	Fix
DAG stuck at `generate_transactions`	Verify `akko-postgresql-data` is `Running`; seed user `alice` exists
Spark job fails with `NoClassDefFoundError: Iceberg`	Custom `akko-spark:2026.03` image not pulled; `bash helm/scripts/build-images.sh`
`MERGE INTO` returns 0 rows	`raw_transactions` is empty — re-run `generate_transactions`
MLflow run not visible	Check MinIO bucket `mlflow-artifacts` exists (created by `minio-init`)
ADEN answers with "no tables"	OpenMetadata ingest has not run yet — trigger `openmetadata_ingest_dag`