Stress-testing supply-chain decisions under disruption
A competition prototype built for the Meta PyTorch OpenEnv Hackathon. SupplyMind explores how planners can turn incomplete public signals into comparable disruption scenarios before committing inventory, routing or supplier capacity.
Prototype materials available on request.
I can walk through the operating problem, the evidence, the decisions made and the limits of the prototype.
Request a case walkthroughSupply-chain planners must act before uncertainty disappears. Signals differ in freshness and reliability, response options carry different costs, and an aggressive policy can protect continuity while creating avoidable inventory or logistics exposure.
The environment structures signals into a risk state, makes evidence weighting and reward assumptions inspectable, and compares candidate responses across historical-crisis patterns and adversarial conditions. Future validation would test calibration error, policy stability and cost-service trade-offs on held-out scenarios; no live operational performance is claimed.
Signal frame
- Input
- Multiple public indicators of disruption
- Engine
- Scenario construction and evidence weighting
- Output
- A structured risk state for policy testing
Response environment
- Input
- Risk state and candidate action
- Engine
- OpenEnv reinforcement-learning environment
- Output
- Comparable consequences and reward signals
Robustness review
- Input
- Historical and adversarial scenarios
- Engine
- Back-testing and reward-design checks
- Output
- Evidence about where a policy breaks
framing
- Disruption scenarios
- Public-signal synthesis
testing
- OpenEnv
- Historical calibration
assurance
- Adversarial reward tests
- Assumption review
Links uncertain signals to the inventory, routing and capacity decisions they may change.
Makes evidence weights, reward choices and response costs open to challenge.
Uses historical calibration and adversarial conditions as research evidence.
Proposed validation: compare calibration, false alarms and policy stability on held-out scenarios.