projects
Practical builds focused on measurable impact.
EdgeStyleGAN: Optimized StyleGAN2 for edge devices using knowledge distillation, pruning, and quantization, achieving 86.3% parameter reduction (28.27M → 3.88M) and 34.9× model size reduction (722.5 MB → 20.7 MB) with QNNPACK quantization for ARM processors.
This project fine-tunes Qwen3-4B-Instruct on the Indic subset of the Aya dataset using Unsloth. The goal is to improve instruction-following and conversational abilities of the model in Indian languages such as Hindi, Bengali, Tamil, Telugu, and others.
Built a hybrid recommender on FAR-Trans (388K transactions, 32K customers, 836 assets) combining Truncated SVD, KNN, content-based and demographic filtering with auto-normalized configurable weights. Addressed cold-start via a MiFID-compliant 25-question risk questionnaire; evaluated with ranking metrics (Precision, Recall, MAP, MRR, nDCG) and business metrics (ROI, Coverage, Novelty) using leave-one-out splitting.
Built a multi-modal character attribute extraction pipeline using CLIP (image) and BART (text), capable of processing 9 structured attributes (age, gender, ethnicity, hairstyle, hair color, eye color, body type, clothing style).
A comprehensive web application that uses RAG (Retrieval-Augmented Generation) to help students upload PDFs, organize by subject/topic, and instantly generate summaries, quizzes, and intelligent Q&A content for fast, effective last-minute studying.