MedGate

Boot-Up Screen for the Application
Node Travsersal is highlighted when agent is navigating relevant context
PDF View
Call logs between systems & LLMS
Special instance (love) - Theme of the hackathon

Inspiration

Healthcare doesn’t have an AI capability problem, it has an information access problem.

The consequences are massive:

Clinicians spend up to ~49% of their time just searching for and documenting patient information— instead of actually treating patients.
~795,000 deaths or permanent disabilities each year in the U.S. from diagnostic errors
Medication errors harm over 1.5 million people every year

A major reason for this is simple: critical information exists, but isn’t accessible in time.

Doctors make decisions using fragmented patient records spread across visits, PDFs, and systems. Signals that could change outcomes are often buried across documents and missed under time pressure.

At the same time, the most powerful AI models could solve this instantly - synthesizing histories, identifying patterns, and grounding decisions in data. But they can’t be used with patient information due to strict privacy constraints like HIPAA.

This creates a fundamental gap:

AI is powerful enough to help
But unusable where it matters most

We built MedGate to bridge that gap, a system designed to support HIPAA-compliant deployments by ensuring all PHI remains on-premise.

What it does

MedGate is a privacy-first AI gateway that enables healthcare systems to safely use models like GPT-4, Claude, and Gemini.

Running on the ASUS GX10, MedGate:

Intercepts user queries
Removes all Protected Health Information (PHI)
Sends only de-identified context to the cloud
Reconstructs responses locally with full identity and source-backed citations

It also acts as an AI agent over clinical data, retrieving information from a structured knowledge graph, querying for additional context, and grounding every answer in real records.

How we built it

We built MedGate as a hybrid system combining local computation with cloud intelligence:

Local Layer ( ASUS GX10)
- PHI de-identification and tokenization
- Knowledge graph storage and traversal
- Ephemeral mapping for re-hydration
- Backend orchestration (FastAPI)
Cloud Layer
- Model-agnostic integration (GPT-4, Claude, Gemini)
- Tool/function calling for structured retrieval
- Multi-step reasoning workflows
Frontend
- Custom chat interface
- Real-time “proof panel” showing redacted vs original queries
- 3D knowledge graph with live traversal visualization
- Clickable citations linked to source documents

We pre-processed hundreds of synthetic clinical documents into a structured knowledge graph to simulate real-world patient histories.

Challenges we ran into

Ensuring reliable de-identification while preserving enough context for accurate reasoning
Designing a secure token mapping and re-hydration system
Supporting multiple model APIs with different tool-calling formats
Balancing performance between local processing and cloud inference
Communicating a complex system clearly within a short demo

Accomplishments that we're proud of

Built a working end-to-end system that enforces privacy at the architectural level
Demonstrated safe integration with multiple frontier AI models
Created a real-time “proof panel” showing that PHI never leaves the device
Developed an interactive knowledge graph that visualizes how information is retrieved
Delivered a clean, intuitive experience for a highly complex backend system