AegisWard

### Inspiration

Hospital ward managers are overwhelmed by the intersection of clinical care and revenue cycle management. We observed that minor documentation gaps like a missing form or a slightly ambiguous note lead to massive, preventable revenue leakage due to payer denials. We were inspired to build AegisWard to automate these administrative burdens, allowing clinical teams to stop fighting billing denials and focus entirely on patient outcomes.

### What it does

AegisWard is a fortified, autonomous clinical operations agent fleet. It operates continuously in the background to:

• Ingest Longitudinal Data: Processes multi-week admissions via an episodic Memory Bank.

• Automate Coding: Assigns accurate ICD-10/CPT/DRG codes grounded in clinical evidence. • Predict Denials: Proactively intercepts documentation gaps based on real-time Medicare NCD/LCD and commercial policy checks. • Enforce Security: Uses Model Armor for strict, inline HIPAA-compliant PHI masking. • Trace Compliance: Provides full W3C OpenTelemetry tracing for audit-ready transparency.

### How we built it

AegisWard is built on a modern, event-driven architecture designed for reliability and security:

• Core: React + Vite + Tailwind CSS for the Ward Command Center. • Orchestration: Node.js/Express API Gateway managing agent DAGs and Pub/Sub event subscriptions. • Intelligence: Gemini 3.5/2.5 via the @google/genai SDK. • Infrastructure: Google Cloud Run, Cloud Pub/Sub, Firestore (episodic store). • Observability & Security: OpenTelemetry for distributed reasoning-chain tracing, and a custom Model Armor layer for inline PHI de-identification.

### Challenges we ran into

• Balancing Autonomy with Trust: Ensuring AI agents made accurate, evidence-based coding decisions without hallucinating clinical context. • PHI Security: Implementing inline, reversible HIPAA-compliant de-identification without breaking the downstream reasoning capabilities of the agents. • Longitudinal Context: Maintaining semantic recall across multi- week admissions while keeping the memory efficient and responsive.

### Accomplishments that we're proud of

• Agent Registry Architecture: Designing a modular, reusable agent ecosystem that can be easily extended to different hospital departments. • Model Armor: Successfully building an inline PHI firewall that effectively balances HIPAA compliance with LLM usability. • End-to-End Tracing: Implementing a complete W3C distributed reasoning-chain trace collector that provides audit-ready cryptographic proof of AI decisions.

### What we learned

We learned that the true power of AI in healthcare isn't just in generating content, but in the intelligent, asynchronous coordination of specialized agents. We discovered that building defensible AI where every cognitive step is traced and audit ready is just as important as the clinical accuracy of the model itself.

### What's next for AegisWard

• Expanded Registry: Creating specialized agents for broader clinical scenarios (e.g., surgical authorization, long-term care management).

• Enhanced Predictive Analytics: Fine-tuning the Denial Predictor Agent for more complex payer-specific prior-authorization rules. • Compliance Dashboard: Developing a dedicated, high-level dashboard for hospital compliance officers to review OTel-derived audit packages in real-time.

Built With

• agent-registry
• antigravity
• billing-automation
• clinical
• docker
• express.js
• firestore
• gemini
• google-cloud
• google-cloud-pub
• google-cloud-run
• google-genai
• healthcare-tech
• node.js
• opentelemetry
• react
• typescript
• vite