About the Project
GeoTwin Sentinel is an observable, agentic digital twin for exploring how regional healthcare infrastructure could respond when climate hazards, infrastructure failures, cyber incidents, and unreliable telemetry occur at the same time.
Most emergency dashboards answer only one question: What is happening? GeoTwin Sentinel goes further by asking:
What could happen next, which intervention appears most effective, why did the agents recommend it, and can the underlying evidence be trusted?
The platform models a synthetic network of five hospitals across the Dallas–Fort Worth region. Users can run three deterministic compound-disruption scenarios:
- Flood Grid Cascade
- Heatwave Ransomware
- Wildfire and Telemetry Tampering
For each scenario, a graph-based digital-twin engine evaluates hospital demand, effective capacity, infrastructure dependencies, cyber-related capacity loss, disruption probability, transfer pressure, regional risk, and resilience.
Four transparent, rule-based agents then collaborate:
- Compound Event Detector identifies interacting physical, operational, and cyber threats.
- Telemetry Integrity Agent evaluates missing, stale, conflicting, or potentially manipulated signals.
- Resilience Planning Agent compares bounded mitigation strategies.
- Response Orchestrator assembles evidence-backed recommendations for human review.
Every agent output includes an explanation, confidence score, evidence references, warnings, execution status, latency, and trace context. Recommendations are never presented as automatically approved actions.
Users can compare five counterfactual interventions against the exact same baseline:
- Backup power
- Surge capacity
- Network segmentation
- Ambulance rerouting
- Telemetry verification
A dedicated Trust and Evidence dashboard exposes evidence quality, telemetry integrity, anomalies, uncertainty, policy checks, agent confidence, and human-review requirements.
The deployed application also includes an interactive SigNoz Simulation Workspace with Overview, Traces, Logs, and Service Map modules. It transforms actual browser-session execution records into trace-shaped observability views while clearly identifying that the hosted demonstration is simulated and not persisted in SigNoz Cloud. The backend is instrumented with OpenTelemetry and can export telemetry to a compatible external collector when one is configured.
GeoTwin Sentinel uses entirely synthetic data and is designed as a research and decision-support prototype. It does not provide clinical advice or autonomously control hospitals, transfers, cybersecurity systems, infrastructure, or emergency services.
Live Demo: https://harshapriyag123-agentic-healthcare.vercel.app
Source Code: https://github.com/harshapriyag123/agentic-healthcare-digital-twi
Inspiration
Healthcare emergencies rarely happen in isolation.
A flood can disrupt power while blocking transport routes and increasing emergency demand. A heatwave can overload hospital capacity while ransomware reduces access to critical systems. A wildfire can create respiratory demand while damaged or manipulated telemetry makes the available evidence unreliable.
Traditional monitoring tools usually display these signals on separate dashboards. They may reveal individual failures, but they do not clearly demonstrate how climate hazards, infrastructure dependencies, cyber conditions, capacity constraints, referral relationships, and evidence quality amplify one another.
GeoTwin Sentinel was inspired by one central question:
How can emergency planners evaluate an AI-assisted recommendation when infrastructure is under pressure and the evidence itself may be incomplete or untrustworthy?
I wanted to build more than a visually impressive risk dashboard. The goal was an inspectable decision-support environment where users could:
- Run a reproducible compound-disruption simulation
- Examine hospital-level and regional consequences
- Understand exactly what every agent considered
- Compare interventions against an identical baseline
- Inspect uncertainty and evidence quality
- Trace the complete workflow instead of trusting a hidden model
- Keep consequential decisions under human control
That combination of digital twins, transparent agents, counterfactual reasoning, geospatial analysis, and observability became GeoTwin Sentinel.
What We Learned
The most important lesson was that observability can become a trust mechanism for agentic systems, not merely a tool for monitoring infrastructure.
A confidence score alone does not prove that a recommendation is reliable. A reviewer must also know:
- Which evidence was available
- Which signals were missing, stale, or conflicting
- Whether telemetry may have been manipulated
- Which rules were triggered
- How long each agent took
- Whether any workflow stage failed
- Whether the recommendation passed policy checks
- Why human review was required
The project also demonstrated why counterfactual interventions must be evaluated against the exact same baseline. If each intervention silently starts from different conditions, the comparison may look persuasive while measuring different problems. GeoTwin Sentinel therefore preserves the baseline and changes only the selected intervention.
Another major lesson was that operational risk and evidence risk must remain separate. A severe disruption supported by reliable telemetry is not equivalent to a severe disruption inferred from incomplete or suspicious data. The interface communicates both dimensions instead of collapsing them into one misleading score.
We also learned that responsible systems must communicate their boundaries directly. GeoTwin Sentinel separates:
- Simulated outcomes
- Observed evidence
- Evidence integrity
- Agent confidence
- Policy compliance
- Recommended actions
- Human authorization
Finally, graceful degradation is essential. An unavailable map tile, missing telemetry signal, failed observability export, or interrupted agent should become a visible and explainable state. It should not silently increase certainty or prevent the complete simulation from producing a reviewable result.
How We Built It
GeoTwin Sentinel is organized as an end-to-end decision-support pipeline:
Scenario configuration → Digital-twin simulation → Agent analysis → Trust evaluation → Counterfactual comparison → Human review → Observability
Digital-Twin Backend
The backend was built with:
- FastAPI
- Python
- Pydantic
- NetworkX
Packaged JSON and GeoJSON datasets represent synthetic hospitals, geographic locations, infrastructure dependencies, referral relationships, hazard exposure, cyber conditions, and telemetry quality.
NetworkX models the healthcare system as a dependency graph. Deterministic simulation rules calculate:
- Projected demand
- Effective hospital capacity
- Capacity shortfall
- Power and infrastructure pressure
- Cyber-related capacity reduction
- Referral and transfer pressure
- Facility disruption probability
- Regional risk
- Regional resilience
The deterministic design makes every scenario reproducible and allows counterfactual interventions to be compared fairly.
Transparent Agent Workflow
The agent layer contains four bounded, inspectable components:
- Compound Event Detector
- Telemetry Integrity Agent
- Resilience Planning Agent
- Response Orchestrator
These agents are transparent rule-based decision components rather than autonomous controllers. Each execution produces structured evidence, confidence, warnings, timing information, policy results, and review status.
Interactive Frontend
The frontend was built with:
- React
- TypeScript
- Vite
- MapLibre GL
The user experience includes:
- Crisis Command Center
- Interactive geospatial hospital map
- Scenario configuration controls
- Hospital-impact tables
- Agent Activity Console
- Counterfactual Explorer
- Trust and Evidence dashboard
- Simulation history
- Architecture explorer
- SigNoz Simulation Workspace
The interface is responsive and includes accessible alternatives to map-based information so users can inspect the same results through structured tables.
Observability
The backend is instrumented with OpenTelemetry across API requests, simulation runs, agent executions, trust calculations, and counterfactual evaluations.
The production application includes a SigNoz-style observability simulation generated from real session execution records. It displays:
- Service health and execution summaries
- Trace-shaped agent workflows
- Correlated structured event logs
- Service and dependency relationships
- Latency and review signals
The application clearly distinguishes this non-persistent demonstration from a live SigNoz deployment. When an external OpenTelemetry-compatible collector is configured, telemetry export operates on a fail-open basis so an unavailable observability service does not stop the simulation.
Quality and Deployment
The project uses:
- pytest
- Vitest
- Ruff
- ESLint
- TypeScript type checking
- Docker
- GitHub Actions
- Render
- Vercel
The delivery workflow includes automated tests, configuration validation, health checks, smoke testing, secret scanning, deterministic fixtures, and a non-root API container.
Challenges We Faced
Maintaining One Consistent State
The hardest technical challenge was ensuring that the map, hospital calculations, agent outputs, trust evaluation, and counterfactual explorer all described the same simulation.
A dashboard can become misleading if each screen independently recomputes values using slightly different assumptions. GeoTwin Sentinel therefore maintains a canonical simulation result and propagates its scenario, facility outcomes, agent records, evidence, and trace context throughout the experience.
Separating Operational Risk From Evidence Risk
Telemetry tampering introduced a more subtle challenge. The platform needed to distinguish between:
- A dangerous event supported by reliable evidence
- A dangerous event supported by incomplete evidence
- A seemingly safe condition created by manipulated telemetry
This required explicit integrity scoring, anomaly detection, evidence lineage, uncertainty indicators, policy checks, warnings, and mandatory review triggers.
Making Agents Inspectable
It was not enough to display polished recommendations. Each recommendation needed to reveal:
- Which rule activated
- Which evidence supported it
- How confident the agent was
- What information was missing
- Which warnings applied
- How long execution took
- Whether the result required human approval
This made agent explainability a first-class data model rather than a paragraph generated after execution.
Building Honest Observability
Observability needed to be part of the product—not a decorative screenshot.
Trace and span identifiers had to connect user actions with simulation stages, agent execution, trust evaluation, counterfactual analysis, and structured events. At the same time, the hosted application needed to remain honest about whether telemetry was simulated, exported, or persisted.
Designing Responsibly for Healthcare
Healthcare resilience is safety-sensitive. The interface had to remain compelling without implying clinical validation, causal certainty, operational authority, or real-world readiness.
For that reason, all data is synthetic, recommendations are bounded, uncertainty is visible, and consequential decisions always remain with authorized human reviewers.
Impact
GeoTwin Sentinel demonstrates how digital twins, geospatial analysis, transparent agents, counterfactual reasoning, trust evaluation, and observability can work together to support more accountable infrastructure-resilience planning.
Its immediate value is as a research, education, and demonstration environment for:
- Emergency-planning teams
- Healthcare-resilience researchers
- Cybersecurity teams
- Public-sector technologists
- Students studying trustworthy AI
- Developers building observable agentic systems
Users can explore questions such as:
- Which hospitals become most vulnerable during a compound disruption?
- How do power, cyber, capacity, transport, and referral dependencies amplify regional risk?
- What happens when an apparently healthy system is supported by unreliable telemetry?
- How does evidence quality affect agent confidence?
- Which bounded intervention performs best under the same baseline?
- Why did an agent make a particular recommendation?
- Where did the workflow slow down or fail?
- Which decision requires human review, and why?
Because the platform uses synthetic data, users can safely experiment with severe failure conditions without exposing patient information or interacting with real healthcare infrastructure.
The broader impact is a design principle:
Agentic systems should not only produce recommendations. They should make their evidence, uncertainty, execution history, limitations, policy status, and human-approval requirements visible.
GeoTwin Sentinel turns that principle into an interactive, testable prototype.
Log in or sign up for Devpost to join the conversation.