Inspiration
In healthcare systems across the globe, Medical Prior Authorization (PA) is notorious for being a slow, paper-heavy bottleneck. Physicians spend over 12 hours a week manually submitting charts, while patients face delayed treatments due to back-and-forth claims friction.
Built PriorPulse AI to prove that modern multi-agent systems powered by Google Cloud Vertex AI and Gemini 2.5 Flash can automate prior authorization in seconds without sacrificing HIPAA security or clinical accuracy.
What it does
PriorPulse AI is an enterprise agent fleet that automates the medical prior authorization process in seconds while enforcing zero-trust HIPAA compliance.
Ingests Multi-Modal Records: Accepts raw EHR clinical notes or uploaded chart PDFs.
Redacts Sensitive PII at the Edge: Uses Google Cloud Data Loss Prevention (DLP) to scrub patient names, SSNs, phone numbers, and dates before transmitting data to any language model.
Extracts Medical Codes & Evaluates Necessity: Gemini 2.5 Flash extracts ICD-10 diagnostic and CPT procedure code pairs, matching them against active insurance policy repositories to verify medical necessity (e.g., conservative physical therapy history).
Enables Physician Sign-Off: Provides a Human-in-the-Loop (HITL) interface for reviewing doctors to approve or deny claims with cryptographic Cloud KMS signatures.
Streams Immutable Audits: Logs every authorization decision, latency metric, and signature directly to Google Cloud BigQuery for compliance auditing.
How we built it
We built PriorPulse AI using a dual-runtime containerized architecture hosted on Google Cloud Run with Artifact Registry:
Frontend & Server Engine: A custom Node.js (server.mjs) Express server driving a responsive, high-contrast SaaS interface (index.html) to avoid third-party CSS overrides.
Agent Execution Fleet: Python execution bridge (runner.py / agent.py) leveraging the Google GenAI SDK (Gemini 2.5 Flash on Vertex AI), @google-cloud/dlp, @google-cloud/bigquery, and pypdf.
Zero-Trust Data Flow: Clinical records are sanitized via edge DLP calls before structured JSON extraction occurs.
Deployment Pipeline: Built into a dual Node 20 / Python 3 Linux container image using Google Cloud Build and deployed with zero-downtime revision management on Cloud Run.
Challenges we ran into
Multi-Runtime Containerization on Cloud Run: Standard Cloud Run buildpacks defaulted to pure Node.js when reading package.json, causing /bin/sh: python3: not found execution errors. We solved this by authoring a custom Dockerfile with explicit Debian Python 3 binaries, system-wide pip package installations (--break-system-packages), and pushing through Google Artifact Registry via Cloud Build.Preserving Clinical Context During PII Redaction: Standard DLP redaction can accidentally mask clinical terms along with patient details. We tuned our Gemini extraction prompts to ignore structural masking tokens while maintaining $100\%$ accuracy when parsing underlying medical necessity criteria.Cross-Origin & Caching Quarks: Overcoming browser-cached Streamlit polling loops after migrating our app stack to pure Node.js Express endpoints required clear API route separation and explicit header handling.
Accomplishments that we're proud of
Sub-1.2s Latency: Achieving an average processing latency of $1.18\text{ seconds}$ from raw document upload to policy rule evaluation.Zero PII Leaks: Demonstrating $100\%$ HIPAA-aligned edge redaction via Google Cloud DLP before LLM prompt execution.Pixel-Perfect SaaS UI: Replacing opinionated frameworks with a lightweight, dark-mode-resilient web workspace that renders cleanly across all reviewer devices.Production-Grade Auditability: Cryptographically signing every physician attestation with Cloud KMS hashes and streaming them live into BigQuery.
What we learned
Gemini 2.5 Flash Excellence: Gemini 2.5 Flash is exceptionally fast and accurate at structured JSON extraction from complex medical text, making it ideal for real-time healthcare workflows.
The Power of Cloud DLP + Vertex AI: Combining edge DLP sanitization with Vertex AI provides a reliable model for building HIPAA-compliant healthcare applications on public cloud infrastructure.
Dual-Runtime Container Optimization: Packaging Node.js for low-latency web serving with Python for specialized AI/ML SDK execution creates a versatile web stack.
What's next for PriorPulse
EHR System Integration: Building native HL7 / FHIR API connectors to pull clinical notes directly from Epic, Cerner, and Athenahealth.
Multi-Document Claims Bundling: Expanding the agent fleet to cross-reference multi-page surgical histories, lab reports, and longitudinal patient charts simultaneously.
Real-Time Payer API Sync: Connecting the policy engine directly to insurance payer clearinghouse APIs for instant, automated claim submission and electronic approval receipt.
Log in or sign up for Devpost to join the conversation.