AISMR
Backend & AI workflows
An AI production workflow with durable jobs and human approvals.
- Python
- FastAPI
- LangGraph
- PostgreSQL
- OpenTelemetry
- The problem
- A useful AI workflow needs more than a generated answer. Work must survive interruptions, connect to tools, and leave decisions with the person running it.
- The workflow
- A creative brief moves through ideation, approval, production, rendering, a second approval, and publishing.
- The foundation
- A FastAPI service, database-backed queue, checkpoints, and workers coordinate the work. Explicit approval gates control the consequential steps.
Engineering notes
- Durable execution
- PostgreSQL-backed jobs, checkpointing, and worker execution preserve workflow progress beyond a single request.
- Controlled actions
- Human approval gates, HMAC-verified webhooks, database row locks, and fail-closed checks constrain model-driven actions.
- Explainable operations
- Request IDs, structured logs, OpenTelemetry traces to Jaeger, health endpoints, tests, and decision records make the system inspectable.