Soujatya Sarkar

Hi, I'm a Computer Science Researcher at Turing working in RLHF and SFT of Gemini (client: Google Deepmind) and also have been appointed as Project Assistant - I at BharatGen (previously BharatGPT) funded by DST and TIH Foundation, IIT Bombay to contribute significantly in building GoI's in-house multilingual and multimodal LLM (especially focussed on supporting all 22 Indic languages) which will be accesible by every citizen of India. Along with that, I am also a Masters by Research student of Intelligent Systems at Indian Institute of Technology, Mandi where, I am presently advised by Prof. Rohit Saluja and am working on advancing seminal architectures using Kolmogorov Arnold Representation Theorem.

I have completed my BTech in Chemical Technology from University of Calcutta in July '23 . Presently, along with my MTech (Research) in Intelligent Systems at School of Computing and Electrical Engineering, IIT Mandi, I am also pursuing a BSc (Hons) in Data Science and Artificial Intelligence from IIT Guwahati to be completed by Aug '28.

My strengths include complete mathematical understanding of NNs, CNNs, RNNs, LSTMs, GRUs, Transformers (including BERT, GPT, LLAMA, Gemma, ViT, DETR, etc.) and GenAI architectures like VAEs, GANs, DDPMs, DDIMs, LDMs. I also have thorough grasp on coding these architectures from scratch in Pytorch. My hobbies include playing around these models, tweaking their losses or normalizing their weights to optimize training and accuracy and also sometimes adding new blocks to them, to create whole new architectures for specific use-cases. For checking out my previous research work, you can scroll down this website or take a look at my Resume/CV.

Presently, I am seeking a researcher/intern role at top corporates in the intersection of large language models, reinforcement learning and computer vision domains.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Github

Research

My research pursuits encompass a broad array of fields, including brain-computer interfaces, human-robot interaction, neuromorphic computing, bioinformatics, psychoinformatics, computer vision, natural language processing, robotics, cheminformatics, and all sorts of applications of machine learning.

Due to severe health hazards arising from skull incisions required for wired BCI/BMIs, scientists are focusing on developing VLSI wireless BCI implants using biomaterials. However, significant challenges, like power efficiency and implant size, persist in creating reliable and efficient wireless BCI implants. With advanced spike sorting techniques, VLSI wireless BCI implants can function within the power and size constraints while maintaining neural spike classification accuracy. This study explores advanced spike sorting techniques to overcome these hurdles and enable VLSI wireless BCI/BMI implants to transmit data efficiently and achieve high accuracy.

Ceramic glaze is an impermeable coating of a vitreous substance fused to a ceramic-ware vessel by firing and is mainly used to colour, decorate and waterproof the item. The present study will mainly focus on the prediction of a scale of addition of the additives (only the binder, i.e., carboxy-methyl-cellulose and water) to adjust the parameters of the Brown Silica Glaze to the production standards, and thus will also improve the quality of the glaze by discontinuing the use of flocculating and deflocculating agents, like, Calcium Chloride (20% solution) and Sodium Hexametaphosphate, which decreases the quality of the glaze. This study has been conducted in partnership with Aditya Birla Insulators, Ltd. using their commercial Brown Silica Glaze for our experimentation purposes.

Miscellaneous

2024 © Soujatya Sarkar | Updated: Dec '24 | Just another barrony.