Gracestack Ai

Inspiration

Healthcare systems forget. Every time a patient arrives, clinicians manually piece together history from fragmented records while the patient waits. We built Gracestack AI to solve this — a multi-agent system with genuine cognitive memory that learns, retains, and reasons about patients over time.

What it does

Gracestack AI is a live multi-agent Clinical Decision Support system. A clinician describes a patient in natural language and within seconds receives:

• Triage assessment (Low/Medium/High/Critical)
• Clinical summary with full patient history
• Drug interaction warnings with confidence scores
• Automatic FHIR R4 journal write
• Gut Feeling anomaly flags

How we built it

Three specialized agents coordinated via A2A protocol:

• Triage Agent (port 10020)
• Memory Agent — Gracestack cognitive core (port 10021)
• FHIR Agent (port 10028)

LLM backbone: Gemini Flash 2.5 Frontend: React | Backend: TypeScript + Rust Infrastructure: Self-hosted VPS Stockholm, Nginx + Docker

Challenges we ran into

Integrating A2A agent discovery with persistent cognitive memory required a custom bridge layer. FHIR R4 normalization from natural language needed careful prompt engineering and clinical validation.

Accomplishments that we're proud of

• Live system processing clinical scenarios in under 3 seconds
• Drug interaction detection at 85% confidence
• 8-memory patient history retrieval across sessions
• Full FHIR journal write on every interaction
• Swedish + English multilingual support
• Built and deployed in one day

What we learned

That the hardest part of healthcare AI is not the intelligence — it is the memory. Most systems are stateless. Gracestack AI proves that cognitive memory transforms a useful tool into a genuine clinical partner.

Built With

• a2a
• anthropic-claude
• docker
• fhir-r4
• gemini-flash-2.5
• hapi
• mcp
• nginx
• node.js
• postgresql
• react
• rust
• typescript