Improving sampling for Diffusion LLMs like LLaDA with Prof Goldblum's group.

• Looking at past predictions for positions

• Improving KV Caching for diffusion LLMs.

Developed novel LLM Eval methods for RAG pipelines at the Machine Learning Center of Excellence.

Worked as an Applied Scientist for 2 yrs at the Search Technology Center India.

• Trained and optimized XLM-R models, cutting L1 latency from 200ms+ to 12ms at 20B+ webpage scale, generating 1M+ new events.

• Applied optimizations including ONNX conversion, model pruning, and Blingfire tokenization, deploying models to serve live production traffic.

• Implemented end‐to‐end data refresh pipeline, resulting in a 1.5M DAU gain

• Reduced hallucinations by 15.7% with SegRank, a Novel LLM Grounding method

• Won the Award of Excellence for Innovation in Microsoft, IDC for my work on the events datapipeline.

• Honourable Mention at the Executive Challenge in the Microsoft Global Hackathon 2024

• Formulated a novel metric CausalCite: a causal formulation of paper citations

• Conducted highly parallelized experiments on a Citations graph with 2B+ edges and 200M+ nodes

• Introduced innovative evaluation metrics with 30.14% better correlation with test-of-time than citations

Published at ACL 2024, Bangkok