🩺 MediMind — Context-Aware Clinical Triage Assistant

AI that thinks in timelines, not snapshots.

🧠 About the Project

MediMind was inspired by a critical gap in existing digital health tools: most symptom checkers treat each patient input as an isolated event. In real clinical settings, diagnosis depends heavily on temporal context — how symptoms evolve over time. A fever today and abdominal pain tomorrow are not independent signals; they form a pattern.

In India, where the doctor-to-patient ratio sits at approximately $1:1500$, the window between symptom onset and professional consultation is where outcomes are often decided. MediMind is designed to operate in that gap — not as a replacement for doctors, but as an intelligent first layer that augments clinical workflows.

A second motivation was emergency response latency. In life-threatening conditions such as cardiac events or strokes, even a few seconds spent waiting on model inference can be dangerous. This led to the design of a non-LLM emergency override — a rule-based detection layer that triggers an SOS flow instantly, completely independent of AI response time.

🏗️ How the Project Was Built

MediMind is implemented as a lightweight, modular full-stack system:

Layer	Role
Flask Backend	Symptom ingestion, event storage, AI inference, report generation
Persistent Event Log	Short-term clinical memory — stores and chains interactions over time
LLM Layer	Structured medical extraction, triage reasoning, SOAP report generation
Frontend	Voice input, image upload, real-time UI state, locked emergency mode

The system follows an event-driven architecture: every interaction is stored, contextualized, and re-used in future reasoning steps rather than discarded. This is what separates MediMind from a standard chatbot.

User Input (voice/text/image)
        │
        ▼
Emergency Override Check (rule-based, instant)
        │
        ├──► 🚨 SOS Flow (if critical phrase detected)
        │
        ▼
Event Log + Timeline Builder
        │
        ▼
LLM Inference (contextualized prompt)
        │
        ▼
SOAP Report Generation + Triage Output

Stack: Python · Flask · LLM Prompt Engineering · Multimodal Input · Event-Driven Architecture · Real-time Frontend

✨ Key Features

🕐 Temporal Clinical Memory — symptoms are tracked across sessions; patterns surface automatically
🚨 Non-LLM Emergency Override — rule-based SOS trigger with $t_{\text{response}} \approx 0$, no model inference required
📋 SOAP Report Generation — structured Subjective / Objective / Assessment / Plan output ready for clinical handoff
🎙️ Multimodal Input — voice, text, and image upload with graceful fallback across model capability tiers
🔒 Locked Emergency UI — frontend enters a restricted state during emergencies, preventing accidental dismissal

📚 What I Learned

Designing context-aware AI systems instead of stateless chat interfaces
Prompt engineering for structured, low-hallucination clinical outputs with consistent SOAP section separation
Handling model capability limitations (text-only vs. multimodal) with graceful fallback logic
Secure deployment practices including environment-based secret management
Translating loosely defined user input into clinically meaningful documentation

$$\text{Strong Architecture} \gg \text{Better Model}$$

This project reinforced a core principle: system design consistently outperformed simply swapping in a more capable model.

⚠️ Challenges Faced

1. Consistent SOAP Report Quality Early versions produced vague, poorly separated clinical sections. Solving this required rethinking both input representation (timeline normalization) and prompting strategy to enforce strict medical documentation standards.

2. API Instability & Provider Switching Rapidly evolving AI APIs and quota exhaustion forced mid-build provider switches, pushing the design toward provider-agnostic inference handling so the system stayed functional regardless of which backend was available.

3. Emergency Detection Architecture The SOS layer required complete separation from the LLM pipeline. Rule-based logic must guarantee near-instant response — model latency is never acceptable in a cardiac event scenario.

🎯 Outcome

MediMind demonstrates how AI can move beyond chat-style interactions into decision-support systems that respect time, context, and safety.

It serves as a proof-of-concept for:

Intelligent triage — prioritizing critical cases before a doctor is involved
Clinical handoff automation — SOAP reports reduce documentation burden on practitioners
Emergency-aware AI interaction — hard separation between AI reasoning and life-critical response

Built to augment clinical workflows, not replace the professionals behind them.

🛠️ Tech Stack

Category	Tools
Backend	Python, Flask
AI/ML	LLM APIs, Prompt Engineering, Multimodal Inference
Frontend	HTML/CSS/JS, Voice API, Real-time UI
Architecture	Event-driven, Provider-agnostic inference
Deployment	Environment-based secret management, GCP