I am a Research Engineer at Google, working on model quality for Gemini's code capabilities in the post-training phase. I design the training recipes, reward architectures, and evaluation frameworks that define how Gemini reasons about, writes, and ships code at scale.
I led the post-training strategy behind Gemini's #1 rank on the WebDev Arena. Core contributor to Gemini 3.0, Gemini 2.5, and Gemini 2.0 — delivering state-of-the-art results across LiveCodeBench Pro, Terminal Bench 2.0, and SWE-bench. My work spans reinforcement learning for code, reward modeling, and long-horizon agentic evaluations.
Previously, I built enterprise LLMs at Capital One (Llama 2, Mixtral, DPO/RLHF), worked with Google DeepMind on multimodal LLMs for document extraction, and published research on multimodal fact-checking (SIGIR — Best Paper Honorable Mention), hate speech detection (EMNLP), and efficient prompt tuning (ACL). I hold an M.S. from Virginia Tech and a B.S. from D.J. Sanghvi College of Engineering.
Highlights
- Core contributor to Gemini 3.0, 2.5, 2.0 — post-training for code, reasoning, and software engineering.
- Pushed Gemini to #1 on WebDev Arena. Multifold improvements on LiveCodeBench Pro, Terminal Bench 2.0, SWE-bench Verified & Pro.
- Best Paper Honorable Mention at SIGIR 2023 for multimodal fact-checking.
- Program Chair, COLING 2025. Reviewer for NAACL, EMNLP, COLING, ACL.
- Google India Scholar. AI Research Fellow at Fellowship.ai.
Selected Papers
Experience
-
ML Research Engineer — Gemini 2024 – presentGoogle · Mountain View
Core contributor to Gemini 3.0, 2.5, and 2.0. Developing post-training recipes (SFT, distillation, RLVR) to push state-of-the-art across code, reasoning, and software engineering benchmarks. Key contributions to #1 on WebDev Arena, multifold improvements on LiveCodeBench Pro, Terminal Bench 2.0, SWE-bench Verified & Pro.
-
Senior Data Scientist 2023 – 2024Capital One · McLean
Built in-house LLMs with Llama 2 and Mixtral — pre-training on enterprise data with Megatron/FSDP, fine-tuning with DPO/RLHF for chat alignment. Implemented KV cache, reduced precision, Multi Query Attention, and Rotary Position Embeddings to optimize training and inference pipelines.
-
Research Scientist Intern — LLMs Sep – Dec 2022Google DeepMind · Seattle
Integrated soft prompt parameters and adapters in a Multimodal LLM for document extraction. Performed parameter-efficient prefix fine-tuning on TPUs. Reduced catastrophic forgetting by 14% through sequential uptraining.
-
Data Science Intern — NLP Jun – Aug 2022Capital One · McLean
Fine-tuned transformer-based language models (RoBERTa, XLNet, T5) on enterprise call transcript data. Improved customer request fulfillment through co-reference resolution, eliminating 70% of false positives with 94% accuracy.
-
Research Scientist — Machine Learning 2020 – 2021Indian Institute of Technology (IIT) · Indore
Developed novel multimodal architectures for sarcasm detection, outperforming benchmarks by 6.14% F1. Proposed efficient self-attention models for code-mixed language. Two papers published at ICONIP and ACL workshops.
-
Machine Learning Engineer 2020Saarthi.ai · Bangalore
Developed deep learning models (BiLSTM, 1-D CNN) for gender identification from audio, achieving 96% accuracy — a 15% improvement over previous approaches.
Education
Service & Honors
- Program Chair — COLING 2025
- Best Paper Honorable Mention — SIGIR 2023
- Reviewer — NAACL SRW 2023, ICON 2023, EMNLP 2022, COLING 2022, ICON · ACL 2021
- AI Research Fellowship — Fellowship.ai, 2020
- Best Research Project — HaXplore, IIT BHU, 2019
- Google India Scholarship — 2017