Hello, I'm JATIN MEHRA

A powerful web content crawler with LLM-powered RAG (Retrieval Augmented Generation) capabilities. CrawlGPT extracts content from URLs, processes it through intelligent summarization, and enables natural language interactions using modern LLM technology.

This Web APP extracts captions from YouTube videos, generates summary, text embeddings, and allows users to search within podcast transcripts. It also refines the context and fact-checks claims using AI models and web crawlers.

Agentic RAG using FastAPI, FAISS, LangChain & Groq — with real-time web validation via Tavily — to answer your PDF-based queries intelligently. Achieved Semantic Similarity (Mean) 0.852 with ~86% evaluation accuracy (threshold 75 semantic similarity scores). Android app using Java.

The smolLM2 135M Ins. MODEL was fine-tuned on the MIT Plagiarism Detection Dataset for improved performance in identifying textual similarities. Achieved 0.96 F1 score, 0.96 recall, and 0.96 precision scores. This model provides binary classification outputs with 1000+ Downloads/Month.

Developed a state-of-the-art AI model for automated essay evaluation as part of the Kaggle AES competition. Achieved 0.79 QWK (Quadratic Weighted Kappa) score and ranked in top 10%. The project aimed to reduce manual grading effort and enhance the feedback process for students and educators.

Convert Podcasts into interesting Blogs using AI, Generate FAQs, Social Media Posts, Newsletter and SEO Elements. This app uses langchain library, llama 4 model, OpenAI whisper-large-v3-turb, Pydantic and fastapi.

Advanced computer vision project for dashcam collision prediction using state-of-the-art VideoMAE-2 architecture. Achieved 11th rank (top 1%) in the prestigious Nexar Dashcam Challenge on Kaggle. Implemented sophisticated video understanding models to predict potential collisions from dashcam footage, leveraging transformer-based video analysis for real-time safety applications in autonomous driving systems.

Advanced computer vision project to detect and classify fake or manipulated scenes in images. Achieved 0.93 AUC as part of a Kaggle competition with top 10% solution. Utilizes deep neural networks and advanced image processing techniques to identify digitally altered content for media authenticity verification.

Machine learning project for predicting blood glucose levels in Type 1 Diabetes patients using the BrisT1D dataset. Achieved 86th rank (top 10%) in Kaggle competition. Implemented time series forecasting models to help patients and healthcare providers better manage diabetes through predictive analytics.

AI-powered application that automatically generates professional resumes tailored to specific job descriptions. Uses natural language processing to analyze job requirements and optimize resume content for better job matching and ATS compatibility.

Advanced multi-agent AI system for conducting comprehensive research on complex topics. Uses collaborative AI agents with specialized roles to gather, analyze, and synthesize information from multiple sources for in-depth research reports and insights.

Developed a deep learning solution for predicting gas turbine electricity output using LSTM neural networks. The model processes time-series data to forecast power generation with high accuracy (RMSE < 370), outperforming traditional prediction methods and reducing the need for manual monitoring.