HeatRelay

The problem

Neighborhood heat response is often coordinated through group chats, spreadsheets, and memory. Small teams must combine alert severity, consent, language, volunteer capacity, cooling-center capacity, mobility needs, cancellations, closures, and outreach status.

What we built

HeatRelay is an operational co-pilot for neighborhood nonprofits, libraries, housing associations, mutual-aid groups, and small local response teams. It turns a verified heat signal into a consent-aware, resource-feasible plan, reconciles it as field conditions change, and stops where a human must decide. The demo is synthetic and replayable: 18 opted-in residents, 5 volunteers, and 2 cooling centers.

How it works

Situation verifies the signal. Needs summarizes only opted-in operational needs. Planning applies deterministic consent, capacity, availability, language, route, and accessibility constraints. Communications prepares multilingual check-ins in a truthful outbox. Watchdog detects cancellations and closures, reconciles the active run, and surfaces evidence when a hard constraint cannot be met. A named coordinator approves or rejects a validated backup route; no agent can approve for a human.

The demo moves from 18/18 coverage to 17/18 after a volunteer cancellation and center closure, surfaces an accessibility decision, and returns to 18/18 only after a documented human approval. Queued is not sent.

Technical implementation

HeatRelay uses the Strands Agents SDK with five focused roles and an agents-as-tools coordinator. The local product uses durable SQLite; the hosted seam supports Amazon Bedrock AgentCore and conditional/versioned DynamoDB. Open-Meteo is the weather adapter and SNS is an opt-in delivery boundary. The repository includes the architecture diagram, evaluator, tests, setup instructions, and reproducible demo assets.

Responsible scope

This is logistics coordination—not medical diagnosis, triage, or an emergency-service replacement. All footage and records are synthetic. The recording does not claim an AWS deployment, live SNS delivery, or measured community outcomes.

Built for the Good Neighbor Agents track in the Agents for Humans Hackathon.

Built With

• amazon-bedrock-agentcore
• aws-sns
• dynamodb
• fastapi
• next.js
• open-meteo
• python
• sqlite
• strands-agents-sdk
• typescript