title: RustVital-AMD
emoji: 🏥
colorFrom: blue
colorTo: purple
sdk: docker
pinned: true
🏥 RustVital‑AMD – Zero‑Trust Medical AI Triage Gateway
Pure Rust • AMD MI300X GPU • Real Web3 Audit Trail
RustVital‑AMD is a production‑grade, zero‑trust medical triage agent that never exposes raw patient data to the AI model. It accepts clinical notes, strips personally identifiable information (PII) before the GPU ever sees them, runs inference on a real Qwen2.5‑7B model backed by an AMD MI300X GPU, and rehydrates the response so the clinician sees a perfectly normal clinical report. Every triage is cryptographically proven and permanently anchored on‑chain via Filecoin CID and Base Sepolia.
🚀 Live Demo: Launch on Hugging Face Spaces
🧬 Key Features
- 🛡️ Zero‑Trust PII Shield – Deterministic, pre‑GPU redaction of names, ages, SSN, MRN, insurance IDs, and more.
- 🧠 Real AMD MI300X Inference – Calls a live vLLM endpoint running Qwen2.5‑7B on ROCm, with graceful CPU fallback.
- 📡 Streaming Agent Observability – WebSocket + SSE endpoint show every pipeline step in real time.
- 📚 PubMed Enrichment – Queries NCBI E‑utilities for relevant biomedical literature after each triage.
- 🩻 DICOM Surrogate Redaction – Demonstrates multi‑modal PII removal from medical imaging headers and burned‑in text.
- 🔐 Cryptographic Audit – A ZK‑simulated proof of correct redaction is generated and verified on every request.
- ⛓️ Blockchain Immutability – Each triage produces a Filecoin CID and an on‑chain transaction on Base Sepolia (via alloy‑rs).
- 🏥 Consortium Attestations – Simulates three independent hospital attestations, creating a decentralised trust model.
- ⚙️ Federated Fine‑Tuning Trigger – Exposes an endpoint that launches a LoRA fine‑tuning job on MI300X, using only redacted data.
- 🎨 Professional Dark UI – Built with HTMX + Alpine.js + Tailwind CSS, served from Axum with no JavaScript build step.
🔄 How It Works
┌───────────────┐ ┌─────────────────────┐ ┌───────────────────────┐
│ Clinician │ │ RustVital-AMD │ │ AMD MI300X GPU │
│ (Browser) │ │ (Axum Server) │ │ (vLLM + ROCm HIP) │
└───────┬───────┘ └──────────┬──────────┘ └───────────┬───────────┘
│ Submit consent form │ │
│────────────────────────►│ │
│ │ (1) PII Shield │
│ │ Redact PHI from note │
│ │ Generate ZK‑style proof │
│ │ │
│ │ (2) Send redacted prompt │
│ │─────────────────────────────►│
│ │ │ Run Qwen2.5‑7B
│ │◄─────────────────────────────│
│ │ (3) Clinical synthesis │
│ │ │
│ │ (4) Enrich with PubMed │
│ │ (5) Compute CID + anchor │
│ │ on Base Sepolia │
│ │ (6) Simulate consortium │
│ │ attestations │
│ │ │
│◄────────────────────────│ (7) Return rehydrated result│
│ View real report │ │
│ (PII never revealed) │ │
🛠️ Technology Stack
| Layer | Technology |
|---|---|
| Language | Rust (edition 2021) |
| Web Framework | Axum + Tower + Tokio |
| Inference | vLLM (OpenAI‑compatible) on AMD ROCm |
| LLM | Qwen2.5‑7B‑Instruct |
| PII Shield | Custom regex engine (deterministic) |
| Crypto | SHA‑256, ECDSA (k256), ZK‑simulated |
| Blockchain | alloy‑rs → Base Sepolia testnet |
| CID | cid + sha2 → Filecoin‑compatible |
| Medical APIs | NCBI E‑utilities (PubMed) |
| UI | HTMX + Alpine.js + Tailwind CSS |
| Observability | WebSocket + Server‑Sent Events |
| Deployment | Docker, Hugging Face Spaces |
📋 Use Cases
🚑 Emergency Room Triage
A clinician pastes a raw ambulance report. The system redacts patient identity, runs the case through the MI300X, and instantly returns a structured clinical assessment with PubMed references – all while maintaining HIPAA compliance and creating an immutable legal audit trail.
🏥 Multi‑Hospital Consortium
Three independent hospitals can run RustVital‑AMD nodes. After each triage, they attest to the correctness of the redaction proof. The on‑chain record shows consensus, giving regulators and insurers cryptographic proof that PII was never exposed.
📊 Continuous Federated Learning
Hospitals trigger fine‑tuning jobs on their own AMD GPUs using only de‑identified logs. The model improves without any raw patient data leaving the gateway. The system logs every training event, creating a fully auditable model lineage on‑chain.
⚙️ Quick Start (Local Development)
Prerequisites
- Rust 1.90+ (stable‑toolchain)
- An AMD MI300X instance running vLLM (or the Hugging Face Space for immediate testing)
- Base Sepolia wallet (optional, for on‑chain transactions)
1. Clone the repository
git clone https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/rustvital-amd
cd rustvital-amd
2. Configure environment
Create a .env file:
VLLM_URL=http://129.212.188.154:8000/v1/chat/completions
PUBMED_EMAIL=you@example.com
PRIVATE_KEY=your_base_sepolia_private_key_hex
PORT=7860
3. Build & Run
cargo build --release
cargo run
The server will start on http://localhost:7860. Open the URL in your browser.
4. Connect to real GPU
If you have a droplet running vLLM, expose it with Cloudflare Tunnel:
cloudflared tunnel --url http://localhost:8000
Then update VLLM_URL with the generated https://*.trycloudflare.com/v1/chat/completions address.
📡 API Endpoints
| Method | Path | Description |
|---|---|---|
GET |
/ |
Serves the interactive UI |
GET |
/health |
Liveness check (healthy) |
GET |
/status |
GPU status, model info, triage count |
POST |
/triage/html |
Form‑encoded triage request (returns HTML) |
POST |
/triage |
JSON triage request (TriageRequest → TriageResponse) |
GET |
/triage/stream |
SSE stream of agent events (Shield → Triage → Audit) |
GET |
/agents/status |
WebSocket for real‑time agent states |
POST |
/dicom/redact |
Upload a DICOM file for metadata / burned‑in text redaction |
GET |
/medical/enrich |
Fetch PubMed evidence for a given clinical note |
POST |
/trigger-federated-tune |
Launch a simulated federated LoRA fine‑tuning job |
GET |
/federation/round |
Show current federated learning round and nodes |
GET |
/dashboard |
HTML fragment of recent triage history |
GET |
/history |
JSON list of all triage records |
☁️ Deployment
The project is containerised with Docker. A pre‑built image is automatically deployed to Hugging Face Spaces on every push.
Dockerfile (production)
FROM rust:1.95-slim-bookworm AS builder
RUN apt-get update && apt-get install -y pkg-config libssl-dev && rm -rf /var/lib/apt/lists/*
WORKDIR /app
COPY Cargo.toml Cargo.lock* ./
COPY src src
COPY static static
COPY contracts contracts
RUN cargo build --release
FROM debian:bookworm-slim
RUN apt-get update && apt-get install -y ca-certificates && rm -rf /var/lib/apt/lists/*
COPY --from=builder /app/target/release/rustvital-amd /usr/local/bin/rustvital-amd
EXPOSE 7860
ENV PORT=7860
CMD ["rustvital-amd"]
Hugging Face Space Secrets
Set these in Settings → Secrets:
| Key | Value |
|---|---|
VLLM_URL |
https://*.trycloudflare.com/v1/chat/completions |
PUBMED_EMAIL |
Your email for NCBI rate limits |
PRIVATE_KEY |
Base Sepolia hex private key (without 0x) |
🏆 Hackathon Tracks
- AI Agents & Agentic Workflows (main)
- Fine‑Tuning on AMD GPUs (LoRA on MI300X)
- Hugging Face Challenge (Space deployed)
- Qwen Challenge (Qwen2.5‑7B via vLLM)
- Build in Public (updates on X/Twitter)
🤝 Contributing
This project was built solo for the AMD Developer Hackathon (May 2026). Contributions are welcome via pull requests. Please discuss major changes first in the Hugging Face Community tab.
📜 License
MIT License – see the LICENSE file (if present) or the Hugging Face Space page for details.
Built with ❤️, Rust, and AMD ROCm for the AMD Developer Hackathon 2026.