ElitorcAI is an advanced multimodal clinical AI assistant that demonstrates real-world agentic workflows using TiDB Serverless for vector search and clinical case analysis. Built for the TiDB AgentX Hackathon, this system showcases multi-step AI agents that go beyond simple RAG to deliver comprehensive clinical decision support.

Multi-Step Agentic Workflow:

flowchart TD
    A[📋 Patient Input<br/>Medical Images + Symptoms] --> B[🔍 Step 1: Medical Image Analysis<br/>Roboflow AI Detection]
    B --> C[🧠 Step 2: Embedding Generation<br/>Cohere Text Embeddings]
    C --> D[🔎 Step 3: Vector Search<br/>TiDB Similarity Search]
    D --> E[⚕️ Step 4: Clinical Reasoning<br/>Groq LLM Synthesis]
    
    F[(🗄️ TiDB Serverless<br/>Vector Database)] -.-> D
    G[📊 MultiCaRe Dataset<br/>200+ Clinical Cases] -.-> F
    
    E --> H[📄 Structured Clinical Report<br/>• Primary Diagnosis<br/>• Similar Cases<br/>• Treatment Plan]
    
    style A fill:#f9f9f9,stroke:#333,stroke-width:2px,color:#000
    style B fill:#f9f9f9,stroke:#333,stroke-width:2px,color:#000
    style C fill:#f9f9f9,stroke:#333,stroke-width:2px,color:#000
    style D fill:#f9f9f9,stroke:#333,stroke-width:2px,color:#000
    style E fill:#f9f9f9,stroke:#333,stroke-width:2px,color:#000
    style F fill:#e6e6e6,stroke:#333,stroke-width:2px,color:#000
    style G fill:#e6e6e6,stroke:#333,stroke-width:2px,color:#000
    style H fill:#f9f9f9,stroke:#333,stroke-width:2px,color:#000

Dr. AMIE (Advanced Multimodal Intelligence for Emergency care) is a sophisticated clinical AI that:

• Analyzes medical images (X-rays, CT scans, MRI) using Roboflow integration
• Searches clinical databases for similar cases using TiDB vector search
• Provides differential diagnoses with confidence levels
• Recommends diagnostic workups and treatment plans
• Ingests new cases into the knowledge base for continuous learning

• Frontend: Next.js with Mastra framework
• Backend: Unified Mastra server with agent orchestration
• Database: TiDB Serverless with vector search capabilities
• LLM: Groq (Llama 3.3 70B)
• Embeddings: Cohere embed-english-v3.0
• Image Analysis: Roboflow API
• Memory: LibSQL with persistent storage

• Node.js 18+
• TiDB Serverless account
• API keys for Groq, Cohere, and Roboflow

git clone https://github.com/N-45div/ElitOrc-AI
npm install
cp env.example .env
# Configure your API keys in .env

npm run seed              # Seed with sample clinical cases
npm run seed:multicare    # Seed with MultiCaRe dataset (with embeddings)
npm run test:tidb         # Test TiDB connection and functionality
npm run reset:tidb        # Reset TiDB tables

# Start Mastra server (port 3001)
npx mastra dev

# In another terminal, start Next.js frontend (port 3000)
npm run dev

Visit http://localhost:3001 for the Mastra playground or http://localhost:3000 for the full chat interface.

Example Queries:

• "A 45-year-old male presents with chest pain, shortness of breath, and fatigue for the past 3 days. He has a history of hypertension and smoking. What should I consider?"
• "Analyze this MRI scan for a patient with lower back pain"
• "Find similar cases for a 67-year-old female with COPD and pneumonia"

1. Patient Query Input → Clinical symptoms, medical history, images
2. TiDB Vector Search → Find similar historical cases using embeddings
3. Image Analysis → Roboflow API processes medical images
4. LLM Clinical Reasoning → Groq analyzes all data for diagnosis
5. Case Documentation → Novel cases ingested back into TiDB
6. Clinical Report → Structured output with recommendations

Patient Input → Embedding Generation → TiDB Vector Search → Similar Cases
     ↓                                                           ↓
Medical Images → Roboflow Analysis → Clinical Context → LLM Analysis
     ↓                                                           ↓
Structured Report ← Clinical Reasoning ← Combined Evidence ← Case History

• Real-World Impact: Actual clinical decision support system
• Multi-Step Agents: Complex workflows beyond simple Q&A
• TiDB Vector Search: Leverages TiDB Serverless vector capabilities
• Continuous Learning: New cases automatically indexed
• Multimodal Integration: Text, images, and structured data

• Clinical case storage with vector embeddings
• Cosine similarity search for case matching
• Scalable cloud-native architecture

• Mastra framework for agent orchestration
• Tool-based architecture with composable functions
• Memory persistence with LibSQL

• Groq: LLM inference with Llama 3.3 70B
• Cohere: High-quality medical text embeddings
• Roboflow: Medical image analysis and detection

• Parquet File Support: Process large medical datasets efficiently
• Batch Processing: Handle thousands of clinical cases with optimized memory usage
• Embedding Generation: Create semantic vectors for similarity search
• Database Migration Scripts: Seamless data import and management

• Multi-modal analysis (text + medical images)
• Evidence-based recommendations with confidence scores
• Similar case retrieval for complex presentations
• Risk stratification based on historical outcomes

• MRI/CT scan analysis with Roboflow integration
• Comparison with similar imaging findings
• Automated reporting with confidence scores

• Case-based learning with similar patient scenarios
• Clinical reasoning development
• Evidence-based medicine integration

• MultiCaRe Dataset Integration: Process large-scale medical case datasets from Zenodo
• Batch Embedding Generation: Efficient processing with Cohere embeddings
• Vector Database Seeding: Populate TiDB with real clinical cases for similarity search

• No PHI stored in vector embeddings
• Anonymized clinical cases only
• HIPAA-compliant architecture ready
• Audit trails for all clinical decisions

This project utilizes the MultiCaRe Dataset for training and evaluation:

• Dataset: MultiCaRe - A Multi-Domain Dataset for Case Report Classification
• Source: Zenodo Repository
• GitHub: MultiCaRe Dataset Repository
• License: Open source dataset for research purposes
• Citation: Nievoff, M., et al. (2024). MultiCaRe: A Multi-Domain Dataset for Case Report Classification

The dataset contains anonymized clinical case reports across multiple medical domains, enabling robust similarity search and clinical decision support capabilities.

• Real-time clinical alerts and notifications
• Integration with EHR systems
• Mobile app for point-of-care use
• Federated learning across healthcare networks

Built for TiDB AgentX Hackathon 2025 | Team: ElitorcAI | License: MIT