Hi there, I'm Prashant 👋

Built an Adaptive RAG application using LangGraph that intelligently decides whether to answer from a local database (FAISS) or live web search. It uses specialized agents to filter out irrelevant documents before generating an answer. The system also double-checks its own work—if it detects a hallucination or an irrelevant answer, it automatically rewrites the query and tries again until it gets it right.

AnalyserGPT – a custom application that analyses the uploaded dataset and answers user queries.
Project is developed using:
• Round Robin Team - to let the agents communicate with each other to achieve final outcome.
• Assistant Agent - to understand the ask, write python code, give final outcome.
• Code Executor Agent - to execute python code.

Implemented two multi-agent patterns to handle complex tasks efficiently.
The Evaluator-Optimizer workflow improves output quality by having agents critique and refine generated answers in a loop.
The Orchestrator-Workers system tackles large problems by breaking them down, delegating sub-tasks to parallel agents, and synthesizing their individual work into a single, comprehensive result.

Developed a MCP based project that demonstrates AI Agent executing tasks in Notion using Notion MCP server.
Tools used within the project:
Autogen - To create Agent & Team.
OpenAI LLM - To provide AI capabilities to the agent.
NGROK - To create Public URL & tunnel b/w Local & Public URL.
Flask - To create a web server
Lovable - To create Frontend

Built an intelligent email assistant that saves time by turning simple subject lines into actionable summaries. Using a custom LangChain tool and secure Gmail integration, the agent autonomously searches through thread history to find the right conversation. It then processes the content to deliver a clear, concise digest of the discussion, highlighting key decision points and necessary action items so you don't have to read every message.

Built an end-to-end Expense Classifier that demonstrates the full lifecycle of custom LLM development. I curated a specialized dataset to fine-tune a Google Gemini model on Vertex AI, tailoring it to recognize specific spending patterns better than generic models. The system is deployed on a scalable endpoint and features a simple Streamlit interface, allowing users to input data and get instant classifications from the custom model.