SentinelTwin

• 

  Secure public entry: product guide first, Cognito login required before any live AI or data action.

• 

• 

• 

• 

  SentinelTwin architecture: authenticated AWS agent loop with CockroachDB as durable operational and vector memory.

SentinelTwin is a human-supervised, multi-hazard disaster-response digital twin whose agents remember.

Emergency teams create valuable knowledge every time they act: what failed, what stayed open, what helped, and what assumptions were made. Most dashboards and AI agents lose that context when a session ends. SentinelTwin connects satellite-derived terrain evidence, bounded simulations, and shared operational memory so the next decision can begin with what the previous response actually learned.

The application supports wildfire, earthquake, compound-event, and agricultural-resilience scenarios. A response loop observes a location, assesses risk, retrieves similar prior situations, simulates response options, and records the decision and outcome for later recall. Human approval and evidence provenance remain visible throughout the workflow.

CockroachDB Cloud is the sole durable memory and operational store. Structured assessments, agent state, simulations, provenance, audits, and fixed-dimension embeddings are kept together. CockroachDB distributed vector indexing supports nearest-neighbor recall beside transactional and spatial data, avoiding a separate vector database and consistency gaps. CockroachDB’s official Agent Skills Repo was executed against the repository: the transaction-design and privilege-hardening skills audited serializable retries, atomic memory writes, vector-query structure, degraded behavior, and least-privilege boundaries. The audit produced bounded full-jitter retries and a regression test. ccloud and Managed MCP remain clearly labeled integration-ready examples and are not claimed as executed tools.

AWS provides the agent environment: AWS Lambda runs the API and agent loop behind API Gateway; Amazon Bedrock supplies planning and reasoning; Amazon S3 stores private, versioned evidence artifacts; Amazon Cognito provides operator authentication; and CloudFront, CloudWatch, X-Ray, Secrets Manager, SQS, and related infrastructure support delivery, monitoring, secret handling, and failure recovery.

The agricultural pathway is evidence-gated. Sentinel-2 features remain evidence, while rainfall deficit, heat anomaly, and irrigation coverage stay explicitly labeled operator assumptions rather than fabricated weather. If durable memory or trusted evidence is unavailable, the system reports a degraded or synthetic state instead of claiming that a write or assessment happened.

The recorded walkthrough demonstrates the end-to-end interaction and the load-bearing memory loop: the first run creates a learned outcome, and the second run retrieves the exact memory identifier and uses it in the next plan. The public deployment now fails closed behind Cognito: guests see the product guide, while authenticated operators can run the live CockroachDB, Bedrock, and S3-backed workflow.

Source code, setup instructions, architecture, security notes, tests, and the MIT license are available in the public repository.

Built With

• amazon-api-gateway
• amazon-bedrock
• amazon-cloudfront
• amazon-cloudwatch
• amazon-cognito
• amazon-web-services
• aws-lambda
• aws-secrets-manager
• aws-sqs
• aws-x-ray
• cockroachdb-agent-skills
• cockroachdb-cloud
• cockroachdb-distributed-vector-indexing