AI That Never Forgets
a Crisis.
ReliefGrid is an autonomous multi-agent disaster response platform powered by CockroachDB Vector Memory and AWS Bedrock. It eliminates operational amnesia by recalling historical disaster decisions to optimize evacuation routing and triage.
Parallel Execution
1024D pgvector Index
Claude 3.5 Sonnet
Real-time Telemetry
📍 Lekki Peninsula Flash Flood
Lat: 6.4698 | Lon: 3.5852 (Lagos Sector)
Ikeja Centre rejected — primary access road submerged during prior storm event.
Cosine Distance: 0.013 | Index: pgvector cosine
Dispatched evacuees to Agege Stadium Hub. Bypassed Ikeja road corridor.
Attribution: Open-Meteo + Vector Memory
Traditional Disaster Response is Severely Hampered by Amnesia.
When disasters hit, decision-makers are flooded with noise. Statistically, up to 70% of response delays stem from fragmented agency data and repeating past shelter or route mistakes.
Disaster Amnesia
Emergency response agencies repeat past failures because operational lessons learned are buried in static post-disaster PDF reports rather than active system memory.
Information Chaos & Fragmented Telemetry
Satellite weather, shelter capacities, road closures, and medical inventory sit in disconnected silos, delaying critical multi-agency consensus during the golden hour.
Hallucinated AI Recommendations
Generic LLM tools lack ground truth verification, producing hallucinated dispatch plans that fail on the ground during active weather and flooding events.
State Loss & System Cascades
Standard stateless AI scripts crash mid-incident, losing progress and forcing responders to restart emergency reasoning from scratch while lives are at stake.
How ReliefGrid Operates Under Fire.
A battle-tested 6-stage autonomous workflow that combines multi-agent consensus, transactional database persistence, and vector memory recall.
Real-Time Incident Ingestion
Incident stream ingested via REST API or operator report, grounded with exact GPS street geocoding.
Coordinator Agent Orchestration
Decomposes complex crisis into concurrent domain mandates for specialized execution.
6 Specialist Agents Parallel Execution
Weather, Infrastructure, Medical, Shelter, Logistics, and Communication agents process domain data simultaneously.
CockroachDB Vector Memory Recall
Performs cosine similarity search over historical decision records to identify past failures and successes.
AI Decision Revision & Evidence Synthesis
AWS Bedrock synthesizes grounded recommendations with explicit data source attribution and confidence scoring.
Actionable Dispatch & Audit Logging
Dispatches emergency units, updates GIS matrices, and logs immutable audit records for after-action review.
6 Specialized AI Agents. One Cohesive Command Network.
Rather than relying on a single monolith prompt, ReliefGrid deploys domain-specific agents executing concurrently under the Coordinator.
ShelterAgent
Evacuation Capacity & Memory-Aware Recall
Manages evacuation shelter allocations, matches displaced population capacity, and recalls past shelter failures.
Queries CockroachDB vector index for prior flood events. Overrides allocations if historical roads were submerged.
OpenStreetMap facility database, displaced population counts, CockroachDB vector decision memory.
Optimal shelter selection recommendations, capacity utilization alerts, access route safety confirmations.
Collective Memory: Turning Disasters into Permanent Intelligence.
Stateless AI models forget what happened yesterday. ReliefGrid embeds 1024-dimensional vectors directly inside CockroachDB, enabling real-time recall that fundamentally alters agent recommendations.
Naïve Shelter Allocation
Evaluates distance only and recommends Ikeja Community Centre. Ignores past flood patterns.
Access road to Ikeja floods 30 minutes after heavy rainfall. Convoy stalls mid-route, stranding 400 evacuees in rising waters.
Memory-Guided Shelter Override
Queries CockroachDB vector index (`DECISION_RECORD` cosine similarity). Recalls prior Ikeja road submergence.
Revises recommendation to Agege Stadium Evacuation Hub. Bypasses submerged Ikeja corridor completely.
7 Standardized Vector Memory Types
Every operational observation, report, and decision is vectorized and queryable across incidents.
Past dispatch overrides & operational decisions
Retrospective findings & bottleneck avoidance
Master emergency response blueprints
Domain reasoning logs & evidence trails
Periodic operational status snapshots
Historical meteorological surge data
Ground responder capacity & damage telemetry
Interactive Real-Time GIS Matrix.
ReliefGrid connects directly to public OpenStreetMap, Open-Meteo, and OSRM protocols to map critical infrastructure with meter-level accuracy.
Lat: 6.5244° N | Lon: 3.3792° E
Capacity: 1,200 evacuees | Status: OPEN
Capacity: 450 evacuees | Status: 85% FULL
Memory Warning: Road Submerged
Built with Best-in-Class Cloud & AI Native Technologies.
No toy scripts or fragile abstractions. ReliefGrid is built on distributed SQL, cloud foundation models, and open GIS standards.
CockroachDB Serverless
Distributed Vector Memory DBHouses 1024D vector embeddings and transactional decision records with pgvector similarity indexing.
AWS Bedrock
Foundation AI ServiceOrchestrates Claude 3.5 Sonnet reasoning & Titan Text Embeddings with enterprise isolation.
FastAPI (Python 3.12)
Asynchronous API GatewayProvides high-throughput async endpoints, request validation, and global correlation middleware.
Next.js 14 + App Router
Frontend Command CenterReact Server Components, dynamic bundle splitting, and strict TypeScript types.
Open-Meteo Weather API
Hydrological & Storm FeedsFetches real-time wind speeds, hourly rainfall levels, and pressure gradients without rate limits.
OpenStreetMap & OSRM
Spatial & Route EngineResolves hospital and shelter coordinates and calculates optimal evacuation drive times.
End-to-End Enterprise Architecture.
Designed for horizontal scalability, zero single point of failure, and complete operational transparency.
Claude 3.5 Sonnet (Reasoning) + Titan Embeddings (1024D vectors)
Distributed SQL + pgvector similarity indexing & serializable transactions
Why CockroachDB is Critical to ReliefGrid.
Single-node SQLite or fragile vector sidecars crumble during real disaster emergencies. CockroachDB provides transactional multi-region vector resilience.
Distributed ACID Consistency
Disaster coordination cannot tolerate split-brain data or partial writes. CockroachDB guarantees strict serializable transaction isolation across nodes.
Native pgvector Embeddings Indexing
Stores 1024D Titan embeddings directly alongside relational operational tables, enabling cosine similarity queries inside standard SQL.
Survivable Multi-Region Infrastructure
During real floods, power grids fail. CockroachDB multi-region deployments survive regional datacenter outages with zero loss of decision state.
Conflict Retry & State Protection
ReliefGrid uses automatic conflict retry handlers to protect agent execution pipelines during heavy concurrent incident ingestion.
Designed for Government & Enterprise Emergency Standards.
Every subsystem is built with strict type safety, end-to-end auditability, and production resilience.
Multi-Agent Consensus Engine
Coordinates 6 domain specialists concurrently under the Coordinator Agent.
CockroachDB Vector Memory
Retrieves prior crisis decisions using pgvector cosine similarity matching.
Watchdog Self-Healing Worker
Detects stalled or timed-out agent tasks and auto-recovers state from checkpoints.
11-Layer Exception Envelope
Standardized error responses with structured error codes, correlation IDs, and trace IDs.
Custom OpenAPI Portal
Interactive enterprise documentation portal built directly inside the web command app.
Ground Truth Telemetry
Integrates real Open-Meteo forecasts, OpenStreetMap facilities, and OSRM routing.
Frequently Asked Technical Questions.
Everything you need to know about the architecture, CockroachDB vector indexing, and AWS Bedrock integration.
Ready to Transform Emergency Response with Memory-Guided AI?
Experience the live Command Center, run simulated flood scenarios, or explore CockroachDB vector recall in action.