A small, executable reference implementation of a supervisor-worker order workflow. It demonstrates typed state, deterministic routing, idempotency, inventory reservation, bounded execution, and an auditable event trail.
This public portfolio version uses in-memory services and synthetic orders. It contains no employer code, client data, model weights, or production claims.
python order_orchestrator.py
python -m unittest -vvalidate → reserve inventory → dispatch → complete
The supervisor stops on rejection and caches terminal results using a payload hash. The worker boundaries are intentionally simple so real database, queue, LLM, and optimization adapters can be introduced without changing state transitions.
- Durable state and transactional outbox
- Kafka-backed workers with retry/dead-letter policies
- Human approval interrupts for low-confidence extraction
- OpenTelemetry traces and per-worker service-level objectives
- Policy-constrained alternative-SKU and fleet optimization tools