Backend & ML builder. CNN, RAG pipeline, deployed APIs. National Top 10 @ TechSprint AI Hack '25. LeetCode 1741.

Durvesh Thorat  • Peerlist

A RAG (Retrieval-Augmented Generation) system that turns PDF documents into a queryable knowledge base — answers grounded strictly in your content, not AI training data. Built with FastAPI, ChromaDB, and Google Gemini. PDFs are chunked into 500-word segments, embedded via gemini-embedding-2, and stored as vectors. Gemini LLM answers using only the top-3 retrieved chunks, eliminating hallucination entirely. Key features: - Auto model fallback chain across Gemini variants for free-tier reliability - Full REST API (upload, query, delete) via FastAPI - CLI mode for direct terminal queries - Dual-interface design for flexibility

Corpus Probe

Handwritten digit recognition CNN built entirely from scratch — no PyTorch, no TensorFlow. Pure NumPy, 4.74M parameters, 99.36% accuracy on MNIST. Training accelerated with CuPy on NVIDIA RTX 3050. Deployed on Hugging Face Spaces with a FastAPI web interface.

Neuro Cipher

An AI-powered Lost & Found platform built for college campuses. Uses Google Gemini for semantic matching between lost and found items — goes beyond keyword search to understand context and intent. Built with React, Firebase, and Google Gemini API. Key features: - Model Cascading failover — auto-switches between Gemini, Groq, and OpenAI on rate limits - Dual media strategy — Cloudinary for hosting with Base64-to-Firestore fallback - Client-side image compression via HTML5 Canvas for mobile optimization - Real-time data flow between React frontend and Firebase NoSQL backend National Top 10 — TechSprint AI Hackathon (GDGPCE × Hack2Skill)