Onward and Upward!

Managed a graduate-level operating systems class of 30 students by facilitating weekly consultation sessions, designing grading rubrics and model solutions, and grading and delivering detailed feedback on assignments.

Under the "Active Portfolio Management with ML and Time Series Forecasting" team.

Ideated, designed, and implemented QBERT, a novel BERT Architecture that leverages Quaternion modules in Attention and Feedforward layers, achieving a 75% reduction in parameters while preserving performance on benchmarks.

Engineered custom frameworks based on Verl and AReaL for agentic reinforcement learning to train LLMs with GRPO and other state-of-the-art RLVR techniques in multi-turn environments (Alfworld, Webshop, etc.).

Conducted post-training (SFT and RLHF) on interface generation tasks using Qwen2.5 Coder language models and synthetic data curated via distributed human feedback, achieving 2x performance gains on interface generation benchmarks while maintaining performance in general benchmarks such as HumanEval and MMLU.

Evaluated LLMs on general benchmarks using the lm-evaluation-harness framework. Contributed to the open-source project by implementing MMLU-Pro and GSM-Plus, two influential benchmarks with 18k+ combined downloads.

Led research and implementation of a novel and efficient way to perform model merging on LoRA modules; authored a research paper on the proposed method, Parameter-Efficient Checkpoint Merging via Metrics-Weighted Averaging.

Led in-depth literature reviews on distributed human feedback and alignment methods, identifying and proposing 5+ novel opportunities for research commercialization.

Spearheaded continued pre-training of Llama 3 70B language models on billions of tokens’ worth of web-scraped corpora using state-of-the-art techniques like Fully Sharded Data Parallel (FSDP) and task arithmetic.