Snehanshu Mukherjee

Research on LLM guided sparsity and learning agents.

• Working on MoE & sparse attention schemes

Worked with Personalisation Team to build in-house recommendation engine.

• Designed & built Axis Bank's first credit card Recommendation System for 4.3Cr customers.
• Developed offer merchant affinity model capability for personalized offers.
• Coded custom matrix operation functions for Spark DataFrames increasing performance of models 3X

Research work on Object Detection and related loss formuations.

• Developed lightweight object detection model (~ 3M params) for Alberta Construction Image Dataset (ACID).
• Achieved 0.71 mAP i.e 7% over baseline mAP, beating YoLoV8 on construction datasets.

Developed multi-agent framework for hyper-personalized e-commerce shopping Implemented few-shot prompting for agents and image search mode. Built a generative agent for customized product descriptions based on user context.

Decoder Based Small language model, coded from scratch. Trained over small story dataset (TinyStories). Finetuned on subset of Wildchat.

C++ library to perform linear algebra on strided tensors CUDA C kernels for basic DL operations @ cuda_sensei Aims to support automatic differentiation and allow neuralnet implementations

Part of AyA expedition'24. Helped integrating SigLip as vision encoder and wrote eval script to calculate BLEU score.

Built Newsense - BERT based text classifier for news articles.

Annual GenAl innovation challenge organised by Rakuten India.

Fixed docstring errors in torch.nn.functional and torch.cuda\, torch.optim\

in proceedings of 2nd International IEEE Applied Sensing Conference (APSCON), 2024