Yash Jain

Building open-source foundation models, owning the full training pipeline - pre-training and post-training.

Building an automatic prompt authoring system. Working on building next set of Office Copilot features.

Master's Thesis: Analysis of mixture of experts on individual, sparsely-annotated and multi-dataset object detection

Graduate Algorithms

Led the development and implementation of a novel ML algorithm that improves speech recognition accuracy by 38.45% compared to existing state-of-the-art, using videos as training data.

Developed a novel pipeline of image difference captioning task for PowerPoint slide data by generating a synthetic dataset in a self-supervised manner, benefitting 4.4. million users. US Patent applied.

• Formulated a novel framework, Group Supervised Learning (GSL), which utilizes synchronous multi-device unsupervised data, extending the principles of contrastive learning to a group setting. • Outperformed supervised and semi-supervised baselines by 0.15 in F-1 score in RealWorld dataset.

1. Developed a novel idea for replacing node features in online social networks by integrating transductive and inductive models (GNNs) for the link prediction tasks. Work published in CIKM, 21 2. Worked on the field of QA systems and Information retrieval with Prof. Soumen Chakrabarti.

Built a QA system to automatically answer user's product-related queries in natural language

1. Built a speech recognition system using RFID stretchable Tattoos placed on the user's face. Tattoos would send their stretch data to a RFID reader which after applying NLP models would detect the word all in real-time. 2. Increasing the range of Near-Field Communication (NFC), we developed a sound theoretical approach along with a working demo that the traditional NFC is not safe anymore due to communication range limitation.

1. Devised an alternative approach for educational assessment of multiple choice questions using Discriminant Index. I also devised the algorithm for carrying out the experiments to further prove the working of the unconventional approach. 2. Created robust scripts for parsing nuclear values from raw datasets to standardize nuclear values like spin, energy level etc. for hundreds of known nuclei.