README.md

---
title: VAJRAM- Multiple Myeloma Agent
emoji: 🔬
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 6.6.0
python_version: 3.10
app_file: app.py
pinned: false
license: apache-2.0
short_description: It is an edge deployable agent for clinical use.
---

# 🔬 VAJRAM: AI-Assisted Multiple Myeloma CDSS

**VAJRAM** is a fully localized, agentic AI Clinical Decision Support System designed for resource-constrained hematology clinics. Running entirely on heavily quantized Vision-Language Models (VLMs), it provides secure, edge-capable analysis without compromising patient data privacy.

### 🚀 Key Technical Novelty: Agentic Mixture of Adapters (MoA)
Traditional multi-agent systems require massive hardware overhead to run multiple distinct LLMs simultaneously. VAJRAM bypasses this bottleneck. The LangGraph orchestrator treats **highly specialized LoRA adapters as its agentic toolset**. By maintaining a single quantized base model in RAM and dynamically hot-swapping specialized adapters for distinct diagnostic tasks (e.g., risk stratification, disease progression analysis), VAJRAM achieves massive multi-agent reasoning on standard edge CPUs. 

### Core Features
* **Vision-RAG Architecture:** Ingests complex oncology guidelines (like ESMO or localized hospital protocols) preserving visual flowchart logic.
* **WSI Patch Analysis:** Integrates multimodal vision encoders to analyze bone marrow biopsy patches.
* **Bring Your Own Data (BYOD):** Agnostic guideline ingestion ensures compliance with enterprise IP and localized standard-of-care practices.

---
### ⚠️ Legal & Clinical Disclaimer
This application is a **Clinical Decision Support System (CDSS)** intended for use by licensed medical professionals. It does not replace professional medical judgment. 
* The AI engine relies on the specific medical guidelines uploaded by the host clinic. 
* The user is solely responsible for ensuring they possess the appropriate enterprise licenses for any copyrighted guidelines (e.g., NCCN, Elsevier/ESMO) uploaded to the vector database.