Track-Reasoning Agents Project Name- TriageMind

Track

Reasoning Agents (Azure AI Foundry)

Project Name

TriageMind

GitHub Username

Sohan-Meghraj

Repository URL

https://github.com/Sohan-Meghraj/TriageMind

Project Description

TriageMind is a customer-support triage agent that reasons in the open, knows its limits, and refuses to break policy. Every complaint runs through six visible reasoning steps — Understand, Classify, Ground, Decide, Draft, Self-check — streamed live to a glass-box UI so you see why, not just the answer.

Instead of the usual answer-or-escalate, it has three honest outcomes: auto-resolve a routine case with a grounded, cited reply; request evidence (a photo) when a damage refund is risky and worth real money; or escalate with a full human briefing when it's unsure, high-severity, or detects a manipulation attempt.

What sets it apart: confidence-gated decisions (auto-resolves only at ≥0.75 with a passing self-check), a self-correction loop, citations to policy docs, hard guardrails including a prompt-injection defense, and an evaluation harness over 27 labeled cases reporting measured accuracy with 0 guardrail violations — visualized in a performance dashboard.

The agent core is built against Microsoft Foundry Agent Service (File Search + function tools via @azure/ai-agents); the client and smoke test are scaffolded. In this build the pipeline runs on a deterministic engine, as live Azure access wasn't available before the deadline — the README documents exactly what's live vs scaffolded.

Demo Video or Screenshots

Live Demo: https://triage-mind.vercel.app/
Screenshots: https://github.com/Sohan-Meghraj/TriageMind/blob/main/docs/screenshot.png

Primary Programming Language

TypeScript/JavaScript

Key Technologies Used

Next.js 16 (App Router), React 19, TypeScript
Tailwind CSS v4, shadcn/ui
Microsoft Foundry Agent Service via @azure/ai-agents + @azure/ai-projects (File Search + function tools — scaffolded)
Server-Sent Events (SSE) streaming
Node/TypeScript evaluation harness; deployed on Vercel

Submission Type

Individual

Team Members

No response

Submission Requirements

• My project meets the track-specific challenge requirements
• My repository includes a comprehensive README.md with setup instructions
• My code does not contain hardcoded API keys or secrets
• I have included demo materials (video or screenshots)
• My project is my own work with proper attribution for any third-party code
• I agree to the Code of Conduct
• I have read and agree to the Disclaimer
• My submission does NOT contain any confidential, proprietary, or sensitive information
• I confirm I have the rights to submit this content and grant the necessary licenses

Quick Setup Summary

Clone the repo, npm install
npm run dev → http://localhost:3000/ (demo runs on the deterministic engine — no keys needed)
npm run eval → 27-case eval report
Foundry (scaffolded): az login + set AZURE_AI_PROJECT_ENDPOINT in .env.local, then npm run smoke Live demo: https://triage-mind.vercel.app/

Technical Highlights

Three-route confidence-gated decisioning (resolve / request-evidence / escalate) — a third "ask for a photo" route for risky refunds
Self-verification + self-correction loop before any reply ships
Input-side prompt-injection guardrail that forces escalation
27-case eval harness, 0 guardrail violations, visualized in a live dashboard
Glass-box UI streaming all 6 reasoning steps over SSE

Challenges & Learnings

Designing one streaming contract so a deterministic engine and the (scaffolded) Microsoft Foundry agent are interchangeable; adding an evidence-request route to balance fraud risk against customer friction; and working within Azure access limits before the deadline.

Contact Information

sohanmeghraj4444@gmail.com

Country/Region

India

This issue is a project submission rather than a scoped change, and it points to the external TriageMind repository. Start with its README.md and the documented setup commands, then inspect docs/screenshot.png and the linked live demo to understand the submitted functionality. No specific implementation task or completion criterion is stated.