HealthTech & Life Sciences VectorPointSubmitted August 1, 2026

Insight RX — an offline multi-agent AI co-pilot that reads chest X-rays and grounds every diagnosis in real medical guidelines

An MCP app on the Model Context Protocol built by VectorPoint at the SRMIST NitroStack × MCP To The Moon hackathon and deployed on NitroStack.

About this project

InsightRX is a multi-agent AI system that reads chest X-rays and drafts a clinical report — entirely offline, with a licensed clinician required to approve every result. What it does: A DICOM chest X-ray plus patient metadata (age, sex, view) enters a LangGraph-orchestrated pipeline. The image is de-identified first — DICOM tag scrubbing and burned-in pixel redaction — and if that fails, the case halts before any AI sees it. A domain-pretrained DenseNet-121 classifies and localizes lung opacities, producing a bounding box and Grad-CAM heatmap so clinicians see where the model is concerned. A diagnosis agent reasons over the findings, an evidence agent retrieves the exact supporting passage from ATS/IDSA guidelines via a signed local FAISS index, and a report agent drafts the write-up. A three-stage verifier firewall then vets it: low-confidence cases abstain, deterministic checks catch inconsistencies, and an independent LLM reviews the reasoning — with calibration (ECE) tracked throughout. Only then does an RBAC-gated clinician approve, edit, or reject, logged to a hash-chained audit trail. Who it's for: Radiologists and hospitals facing imaging backlogs and burnout — especially resource-constrained or rural facilities without reliable cloud access, who can't risk sending scans to third-party APIs. What makes it special: Everything runs locally — Llama 3.1 / Qwen2.5 via Ollama, 4-bit quantized — so no patient data leaves the hospital's hardware. PHI is stripped with Presidio. Every diagnosis is grounded in a cited guideline, and every report clears an independent verification firewall before a human signs off. Nothing is fully autonomous, by design. Future prospects: Monetization follows a compliant per-study SaaS and site-license model, priced against the radiologist backlog it clears, and expands via add-on pathology modules (fracture, nodule, cardiomegaly) and a validated-evidence tier hospitals pay for to support audit and regulatory submissions

HealthTech & Life Sciences track

Design AI-powered solutions for healthcare, diagnostics, patient care, medical research, and life sciences.

Live MCP appSource codeLinkedInLinkedInLinkedInLinkedIn

Team VectorPoint

  • Souryaneel PalLead

  • SURASHRI BANERJEE

  • Parth Sachin Bhardwaj

  • Divyanka Shah

← All SRMIST projects

Frequently asked questions

What does Insight RX — an offline multi-agent AI co-pilot that reads chest X-rays and grounds every diagnosis in real medical guidelines do?
InsightRX is a multi-agent AI system that reads chest X-rays and drafts a clinical report — entirely offline, with a licensed clinician required to approve every result. What it does: A DICOM chest X-ray plus patient metadata (age, sex, view) enters a LangGraph-orchestrated pipeline. The image is de-identified first — DICOM tag scrubbing and burned-in pixel redaction — and if that fails, the case halts before any AI sees it. A domain-pretrained DenseNet-121 classifies and localizes lung opacities, producing a bounding box and Grad-CAM heatmap so clinicians see where the model is concerned. A diagnosis agent reasons over the findings, an evidence agent retrieves the exact supporting passage from ATS/IDSA guidelines via a signed local FAISS index, and a report agent drafts the write-up. A three-stage verifier firewall then vets it: low-confidence cases abstain, deterministic checks catch inconsistencies, and an independent LLM reviews the reasoning — with calibration (ECE) tracked throughout. Only then does an RBAC-gated clinician approve, edit, or reject, logged to a hash-chained audit trail. Who it's for: Radiologists and hospitals facing imaging backlogs and burnout — especially resource-constrained or rural facilities without reliable cloud access, who can't risk sending scans to third-party APIs. What makes it special: Everything runs locally — Llama 3.1 / Qwen2.5 via Ollama, 4-bit quantized — so no patient data leaves the hospital's hardware. PHI is stripped with Presidio. Every diagnosis is grounded in a cited guideline, and every report clears an independent verification firewall before a human signs off. Nothing is fully autonomous, by design. Future prospects: Monetization follows a compliant per-study SaaS and site-license model, priced against the radiologist backlog it clears, and expands via add-on pathology modules (fracture, nodule, cardiomegaly) and a validated-evidence tier hospitals pay for to support audit and regulatory submissions
Who built Insight RX — an offline multi-agent AI co-pilot that reads chest X-rays and grounds every diagnosis in real medical guidelines?
Insight RX — an offline multi-agent AI co-pilot that reads chest X-rays and grounds every diagnosis in real medical guidelines was built by team VectorPoint at the SRMIST NitroStack × MCP To The Moon hackathon, in the HealthTech & Life Sciences track.
What is an MCP app and how is it built?
An MCP app is an application built on the Model Context Protocol — an open standard that lets AI agents connect to tools, data, and APIs. This project exposes MCP tools and resources that agentic AI systems can call. It was built and deployed on NitroStack, the full-stack platform for shipping MCP apps and servers.