Projects

Developed an advanced NLP system to automatically identify and classify mathematical misconceptions from students' written explanations. Built a sophisticated ensemble of 6 large language models (7B-14B parameters) combining LoRA, QLoRA, and full fine-tuning strategies. Achieved 0.947 MAP@3 score, ranking 45th out of 2,500+ participants.

Advanced computer vision project for dashcam collision prediction using state-of-the-art VideoMAE-2 architecture. 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.

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. The project aimed to reduce manual grading effort and enhance the feedback process for students and educators.

Advanced computer vision project to detect and classify fake or manipulated scenes in images. 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. Implemented time series forecasting models to help patients and healthcare providers better manage diabetes through predictive analytics.

AI-powered web content crawler with advanced LLM-powered RAG (Retrieval Augmented Generation) capabilities. Intelligently extracts content from URLs, processes it through sophisticated summarization algorithms, and enables natural language interactions using cutting-edge LLM technology for enhanced information retrieval.

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

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.

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. Provides binary classification outputs.

Convert podcasts into engaging blogs using AI. Generate FAQs, social media posts, newsletters, and SEO elements. Uses LangChain, Llama 4, OpenAI Whisper-large-v3-turbo, Pydantic, and FastAPI.

Web app that extracts captions from YouTube videos, generates summaries and text embeddings, and allows users to search within podcast transcripts. Also refines context and fact-checks claims using AI models and web crawlers.

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.

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.