projects

EdgeStyleGAN

86.3% fewer params34.9× smaller

StyleGAN2 optimized for edge devices with knowledge distillation, pruning, and quantization — 28.27M parameters down to 3.88M, and 722.5 MB down to 20.7 MB via QNNPACK quantization for ARM processors.

code

Aya-Indic-Qwen3-4B

Qwen3-4BUnsloth

Qwen3-4B-Instruct fine-tuned on the Indic subset of the Aya dataset, improving instruction-following and conversation in Hindi, Bengali, Tamil, Telugu, and other Indian languages.

code

Financial Asset Recommendation

388K txns32K customers

A hybrid recommender over FAR-Trans combining truncated SVD, KNN, content-based and demographic filtering. Cold-start handled by a MiFID-compliant risk questionnaire; evaluated on both ranking metrics (MAP, MRR, nDCG) and business metrics (ROI, coverage, novelty).

code

Who's That Character

CLIP + BART9 attributes

A multi-modal attribute extraction pipeline pairing CLIP for images with BART for text, resolving nine structured attributes from age and gender through hairstyle, eye color, and clothing style.

code

Cram AI

RAG

A study tool that ingests PDFs, organizes them by subject and topic, and generates summaries, quizzes, and Q&A on demand for last-minute revision.

code