|

Processed and filtered 5.24M+ crash records using PySpark, Hadoop, SQL, and MLflow; implemented clustered data processing on distributed clusters handling 500K+ rows per batch, reducing analysis time by 60%. Trained an ensemble of three classifiers (XGBoost, Random Forest, LightGBM) with custom preprocessing pipelines for mixed-type features, achieving 93.4% accuracy and 0.93 weighted F1-score via memory-efficient sparse matrix operations, delivering a 5.2% improvement over a linear baseline after iterative optimization.

Developed a 2D upper-body virtual try-on pipeline using diffusion-based inpainting with Stable Diffusion, SAM (Segment Anything Model) for precise garment masking, and advanced image processing techniques. Addressed the critical challenge of online fashion returns (30-40%) by enabling customers to virtually try on clothing. Implemented pose keypoint alignment, thin-plate-spline warping, and ControlNet for structural guidance to achieve realistic garment fitting while preserving identity, pose, and background.

Developed an AI-driven border surveillance system using thermal and night vision with modified Faster R-CNN architecture.

Built a comprehensive disease prediction system using machine learning on medical symptoms dataset. Achieved 97.3% accuracy with multiple classification models deployed on AWS Cloud.

Developed a multi-headed CNN system for real-time sign language recognition with 96% accuracy. Integrated Google F5 TLAN for NLP, reducing translation lag by 95% and achieving 2nd Runner-Up at IIT Bombay.

Built a small language model using PyTorch, trained and tested on the Encyclopedia Britannica and TinyStories datasets. Used a BPE tokenizer, efficient binary storage, a 6-layer Transformer with multi-head attention, and mixed precision training on M3 Pro GPU. Achieved stable training with AdamW and cosine annealing, generating coherent text outputs.

Developed a convolutional autoencoder for image reconstruction, achieving 0.92 SSIM and 38.5 dB PSNR. Implemented multi-stage convolutional and upsampling layers, reducing dimensionality by 75% while preserving 90% visual information. Conducted comparative analysis across epochs, optimizing reconstruction error from 0.0156 to 0.0021.

A simple repository that shows the process of building an MCP server and using Claude Desktop as a client. Features a Travel Desk system to handle employee travel requests, approvals, and history tracking — all accessible directly from Claude. Demonstrates how to modify the contents to develop specific MCP use cases.

Using the advanced Llama 2 7B Chat model by Meta, this project offers a seamless experience for generating high-quality blogs with just a few clicks. Features AI-powered blog generation, customizable writing styles (Fun, General, Professional), word count specification, and a user-friendly Streamlit-based web interface.

Comprehensive analysis of online retail data demonstrating how Fireducks significantly speeds up data processing and analysis compared to traditional methods. Showcases performance improvements in data manipulation, aggregation, and visualization for retail analytics.