Portfolio - Pannaga Sreya U

FinBot – Role-Based AI Assistant (Advanced RAG)

Domain/Function: RAG Based Chatbot

Built and deployed a secure RAG pipeline using Docling, hierarchical chunking, Qdrant, and Groq LLMs. Added RBAC, LangChain/LangGraph orchestration, LangSmith observability, guardrails, and RAGAs evaluation (~90% faithfulness), deployed on Azure AI Foundry.

View Project
on GitHUB

FinBot – Role-Based AI Assistant (Advanced RAG)

Logged in as 'finance' role. Unrelated queries prompt for relevance, marketing questions are denied due to role restrictions, and finance‑related queries like revenue return proper answers with secure role-based access

The RAGAs evaluation says that the answer relevance is 90% and the context recall is 100%

Project Details

•Designed and deployed an end-to-end RAG pipeline using Docling, hierarchical chunking, Qdrant, and LLMs (Groq).
•Implemented role-based access control (RBAC) ensuring secure and governed data retrieval.
•Built LLMOps workflows with LangChain and LangGraph for multi-agent orchestration.
•Integrated observability and tracing using LangSmith for monitoring model behavior.
•Developed guardrails for prompt injection detection, safe responses, and query validation.
•Evaluated system performance using RAGAs (~90% answer faithfulness).
•Deployed on Azure AI Foundry, enabling scalable AI application hosting.

Document Classifier

Domain/Function: Multimodal AI

A document classification system that uses Regex, OCR and Large Language Models (LLM) to automatically categorize medical and administrative documents.

View Project
on GitHUB

This is the UI for the project. We upload/ drag and drop images and classify. Next.js(Typescript) is used here. Deployed on Vercel

1. Prescription took 30secs - LLM used 2. Patient bills took 0.18secs - File name regex 3. KYC document took 5secs - OCR

Project Details

This project classifies uploaded documents into the following categories:

Patient Bills
Claim Forms
KYC Documents
Medical Reports
Prescriptions

Features:
Hybrid Classification Pipeline:
First tries file match using Regex
Then, it tries rule-based keyword matching using OCR
Falls back to LLM classification if OCR cannot determine the category
Batch Processing: Upload multiple documents at once
Drag & Drop Interface: Easy-to-use web interface
Token Usage Tracking: LangSmith integration for monitoring
AWS Deployment

Tech Stack:
Frontend - Next.js, TypeScript, Axios
Backend - FastAPI, Python
OCR - Tesseract OCR (lightweight, ideal for AWS free tier)
LLM - Google Gemini (gemma-4-31b-it)
Tracing - LangSmith

Project video

AI Job Recommendation Agent | Microsoft Foundry, MCP, FastAPI

Domain/Function: Agent in Microsoft Foundry

Built a Job Search Agent using Microsoft Foundry, which uses WebSearch, MCP Server, RAG tools. Takes the skills from resume, finds skill gap considering github repo and skills and recommends relevant jobs

AI Job Recommendation Agent | Microsoft Foundry, MCP, FastAPI

This is a sample of how the agent looks in foundry. We will be able to see the tools, LLM, propmts we used and our question and answer in the output format we mentioned

This image explainsthe flow of how agent works

Project Details

•Recommend jobs based on user role, experience, and location
•Implemented RAG (Retrieval-Augmented Generation) to ingest and query user resumes for personalized job matching
•Built and integrated a custom MCP (Model Context Protocol) server to extract in-demand skills from GitHub
repositories
•Enabled dynamic tool invocation, allowing the agent to use MCP tools for real-time skill analysis and
recommendations
•Performed skill-gap analysis by comparing resume data with industry trends
•Developed backend services using FastAPI and exposed endpoints for tool discovery and execution

CodePilot – Multi-Agent AI Coding Platform

Domain/Function: AI

Codepilot is a multi agent AI coding platform that polls GitHub issues, classifies them into reusable skills, explores the repository, prepares implementation workflows, verifies work in a local sandbox, and asks for human approval before opening a pull request.

View Project
on GitHUB

CodePilot – Multi-Agent AI Coding Platform

PR github

TUI

Project Details

Codepilot polls GitHub issues, classifies them into reusable skills, explores the repository, prepares implementation workflows, verifies work in a local sandbox, and asks for human approval before opening a pull request.

This version uses:
Textual for the TUI
LangGraph for task stages and state transitions
Agents-compatible agent boundaries and prompts
LangChain GitHub Toolkit for GitHub operations
Google Gemini via langchain-google-genai
Local episodic memory under .codepilot/memory

Hello, I am

Pannaga Sreya U

About

Key

My

FinBot – Role-Based AI Assistant (Advanced RAG)

Document Classifier

AI Job Recommendation Agent | Microsoft Foundry, MCP, FastAPI

CodePilot – Multi-Agent AI Coding Platform

My

Awards

Contact Me

Email

Github

Linkedin

Call

Pannaga Sreya U

Download Resume