Human approval for high-risk actions
The agent can recommend an action, but a person remains the final gate before execution.
Case study · AI + backend
Finance-Autopilot
An AI-assisted financial workflow designed to make useful recommendations without treating financial actions as a black box.
System architecture
A request path for the dashboard, an asynchronous path for jobs, and an explicit approval checkpoint.
Next.js dashboard
→
FastAPI service
→
Redis / ARQ workers
→
Human approval
→
Plaid + audit log
How it works
Users view connected financial information in a Next.js dashboard. The application talks to a FastAPI backend that coordinates data access, proposed actions, and audit events.
Work that should not block a request is queued through Redis and processed by ARQ workers. The agent's output is treated as a proposal; a human approval step is required before a high-risk action can proceed.
Technology stack
Next.jsTypeScriptPythonFastAPIPlaidRedisARQPostgreSQLLLM APIs
Engineering decisions
Idempotent background jobs
Retry-safe job handling prevents a transient failure from creating duplicate financial actions.
Retry queue with Redis + ARQ
Long-running work is decoupled from the request path so the API stays responsive and failures can be retried.
Hash-chained audit log
Each audit entry is linked to the prior entry, producing an inspectable history of proposed and approved actions.