MedIntel

Inspiration

Healthcare today is fragmented. People rely on unverified internet advice for everyday symptoms, doctors lose context between visits, and public health teams react late because signals like pharmacy sales and air quality data remain disconnected.
We were inspired to close this gap by building a system that unifies personal health guidance, clinical continuity, and public health foresight — so that every individual, clinician, and community share the same intelligent context.

What it does

MedIntel is an agentic AI health companion that bridges two levels of healthcare:

AI Medical Chatbot — Gives evidence-based answers using Agentic RAG (Retrieve–Act–Generate). It integrates WHO data to provide accurate and safe responses.
Personal Health Record Summaries — Stores user data (with consent) and generates PDFs of medical history for analysis by doctors.

Together, these modules transform healthcare from reactive to proactive, ensuring local, trustworthy, and continuous health insights.

How we built it

We built MedIntel using ngrok for API tunneling and n8n (locally and via Docker) as the orchestration engine.

Knowledge Layer: Documents (WHO, CDC, Harrison’s Medicine) are fetched and chunked hierarchically (1,200 characters with 200 overlap).
Embedding & Search: Google Gemini embeddings are stored in a Vector Store for efficient retrieval.
Agent Flow: Input → Webhook → Memory → AI Agent (Gemini Model + tools: medical_kb.search, env.aqi) → JSON validation → Webhook Response.

This modular Agentic RAG architecture enables safe, explainable, and locally relevant AI health guidance.

Challenges we ran into

Maintaining safety and accuracy while ensuring responsiveness.
Cleaning and normalizing diverse medical PDFs for consistent embeddings.
Enforcing structured JSON outputs and graceful fallbacks in n8n workflows.
Balancing latency vs. reliability for real-time contextual retrieval.

Accomplishments that we're proud of

Implemented a fully functional Agentic RAG chatbot with deterministic citations.
Designed a city-scale outbreak predictor that improves continuously with anonymized user data.
Achieved end-to-end observability with per-node latency tracking and reproducible runs.

What we learned

Safety > fluency: Reliable citations build user trust more than eloquent text.
Agentic RAG ≫ plain LLM — deliberate tool use improves accuracy and accountability.
The future of health AI lies in context fusion: linking patient, clinician, and community data.
Building production-grade AI in low-code orchestration tools (like n8n) requires precise design and validation.

What's next for MedIntel

Neighborhood-Aware Outbreak Prediction —
Extend the city model into a spatio-temporal GNN + LSTM system that learns how diseases spread across nearby regions using mobility data and anomaly fusion.

Key Note: The outbreak model becomes smarter with every user (after full consent). Your data — completely anonymized — helps train a system that can keep your community safer.

Personalized Doctor Interaction —
Enable secure doctor–patient chats within MedIntel, complete with AI-summarized records, triage assistance, and appointment workflows.

These upgrades will turn MedIntel into a living health platform — sensing risks early, guiding individuals responsibly, and helping authorities act before crises emerge.