2026
ReFlow
Recovery agent that replans disrupted container haulage pickups, with a dispatcher approving each change
- Recovery options
- 4
- Role views
- 3
- Approves each change
- Human
- Python
- FastAPI
- Next.js
- Tailwind CSS
- Groq
Summary
ReFlow is a working prototype built for PSA Code Sprint. When a planned container pickup becomes infeasible, it evaluates whether to wait, reassign, swap or reschedule the affected jobs, asks a dispatcher to approve anything that changes a real assignment, executes the change, and keeps replanning until the fleet plan is stable again. It runs against an in-memory simulator, not live terminal systems.
Problem
One missed or delayed pickup cascades: the truck, the driver and every later job on that plan are affected, and dispatchers have to untangle it by hand under time pressure.
Approach
A FastAPI backend keeps a strict line between deterministic tools (rules engine, optimiser, routing, state changes with an audit trail) and LLM-backed agents, which are the only code allowed to call a model and fall back gracefully without one. A Next.js front end gives dispatchers, managers and drivers their own views.