Projects

CourseTA

Agentic-System Langgraph RAG LLMs Human-In-The-Loop Async Reflection-agent NLP

An agent-based educational system using FastAPI, LangChain, and LangGraph in a microservice architecture. Implemented real-time async endpoints, integrated Whisper for audio/video transcription, and PyMuPDF for PDF parsing. Built a RAG-based QA pipeline with embedding-based retrieval, and AI agents for question generation, summarization, and feedback refinement. Optimized for scalable human-in-the-loop workflows and efficient content transformation across diverse formats.

Demo Repo

Group Activity Recognition

Paper-Implementation CVPR-2016 Distributed-Training Torch Two-Stage Arch Spatial-Temporal Multi-Person-Recognition Computer Vision

A modern implementation implemented a Hierarchical Deep Temporal Model for Group Activity Recognition, based on the CVPR 2016 paper. Achieved 93% accuracy using a two-stage LSTM architecture to recognize multi-person activities. Conducted ablation studies to evaluate the contributions of various model components and compared performance against 8 baseline models.

Repo Arxiv

Relational Group Activity Recognition

Paper-Implementation ECCV 2018 Distributed-Training Torch GNN CNN LSTM Graph Attention Spatial-Temporal Computer Vision

Implemented a Hierarchical Relational Network architecture for Group Activity Recognition and Retrieval, based on the ECCV 2018 paper. This implementation models inter-person relations and hierarchical temporal dynamics using Relational Layer (Graph Relational Layer). Ablation experiments were conducted to assess the impact of relation modeling, attention mechanisms (Graph Attention Operator), and hierarchical layers.

Repo Arxiv

ImgCap

CNN RNN Beam-Search Attention mechanism NLP Torch Teacher Forcing

ImgCap is an image captioning model designed to automatically generate descriptive captions for images. It has two versions one based on a CNN-LSTM architecture, and the other an enhanced version that integrates a CNN-LSTM with an Attention mechanism. To further enhance the quality and coherence of generated captions, beam search is implemented during the inference stage.

Repo Space

Credit Card Fraud Detection

Imbalanced-Data Smote Focal-Loss Random-Forest Voting-cls EDA

Credit card fraud detection project using a highly unbalanced Kaggle dataset of 284,807 transactions with only 492 frauds. To address the imbalance the project implements a voting classifier and a neural network with focal loss in PyTorch achieving an F1 score of 86% and a PR AUC of 85%

Repo Report

Tennis Analysis System

Object-Detection Object-Tracking YOLOv5 CNN Ultralytics Torch Computer Vision

Tennis-Analysis-System is a system that analyzes tennis matches and provides player statistics and mini court visualizations.

Demo Repo