Fraud is the highest-volume financial services workload on the Databricks Platform, and it is the one that punishes a split architecture hardest: real-time scoring sits in one system, ninety days of history in another, and the investigator who has to connect them gets neither governance nor speed. Put an AI agent into that picture and you inherit a second problem, which is proving the model only ever saw what the analyst was allowed to see.
In this hands-on workshop you will build the whole flow yourself on AWS, in a private catalog that is yours alone. Governance comes first: you will work with a synthetic card-transaction feed in Unity Catalog, watch card numbers and cardholder names come back masked by a column mask evaluated against your identity, and see that the mask is enforced at the table for every tool that reads the data, not in any one query or application. You will stand up an AI/BI Genie space over those governed tables, ask investigator questions in plain English, and read the SQL Genie writes for you. You will query real-time scoring and fraud case management data in Lakebase, the managed Postgres service on the lakehouse, and join it to your Delta alerts in a single statement with no pipeline in between. Then you will put a foundation model in front of all of it: a notebook calls the Genie API as you, hands the governed rows to Amazon Bedrock, and gets back an analyst's case narrative. The mask you saw in the first lab holds through the SQL editor, through Genie, and through the API call, so the model never sees a card number and nobody wrote a separate privacy policy for the AI path. We close with where this goes in production: Amazon Bedrock AgentCore as the agent runtime and Unity AI Gateway for governing every call an agent makes back into your data.
Everything runs on Databricks Serverless in a workspace we provision for you. There are no clusters to size and nothing to install. You sign in and start building.
What you will build:
- Unity Catalog: card-transaction tables with column masking for PII that enforces privacy at the source for every downstream tool
- AI/BI Genie: a natural-language space where investigators ask questions in plain English and audit the underlying SQL
- Lakebase: a single-statement join connecting real-time scoring and case management records with your Delta history on Amazon S3
- Amazon Bedrock: a complete investigation notebook that uses the Genie API to generate case narratives from governed data
Who should attend:
Data and analytics leaders, platform and data engineering teams, and architects responsible for fraud, risk, or governed AI. The session is technical, level 200 to 300. Familiarity with SQL and Python helps, but no lab requires writing code from scratch.