Dev Logs / AeroPulse

AeroPulse

A review desk for a synthetic 100,000-row transaction ledger. Charts show the risk. A person approves, escalates, or dismisses each flag.

The operational problem

A control team cannot read 100,000 payments one by one, and it cannot treat every system flag as fraud. AeroPulse is the triage surface. The Streamlit dashboard shows volume, flagged risk, anomaly rate, and the pending queue. AuditAgent, a FastMCP server, exposes the same ledger to an IDE as read-only tools: schema, SQL, anomaly rules, and an audit memo.

Technical highlights

  • SQLite ledger with WAL. Review columns (audit_status, notes, timestamp) are added at startup if the generator did not create them
  • Human decisions are Approve, Escalate to fraud, or Dismiss. Dismiss clears the system flag. The row stays in the table
  • Analyst questions go to Gemini or OpenAI when a key is present, and to fixed SQL templates when it is not
  • Generated SQL must be a single SELECT. Results stop at 200 rows. The agent also sets SQLite query_only
  • Anomaly rules are explicit: negative amounts, users two standard deviations above the peer count, and high-value wires or vendor payments booked overnight

Business impact

The generator marks about 1% of rows as negative amounts and about 0.5% as amount spikes. Reviewers work the flagged subset in amount order and download a markdown memo that ranks escalations first. The model is only allowed to propose SQL. It does not close a case.

The ledger is synthetic. Review notes in a hosted copy reset when the app restarts, because the database is restored from the repo. The full write-up documents each AuditAgent tool and the HITL write path.