Nov 2024 – Jul 2025 · 8 mos · Zurich, Switzerland

• Gemini reasoning and post-training.
• Post-training and RL scaling for Gemini 2.5 Pro thinking model.
• https://deepmind.google/discover/blog/gemini-achieves-gold-level-performance-at-the-international-collegiate-programming-contest-world-finals/

Apr 2023 – Nov 2024 · 1 yr 7 mos

• Gemini reasoning and post-training
• Core contributor in the post-training recipes for Gemini 1.0 and Gemini 1.5 family of models.
• Reasoning: Developed novel training recipes in Gemini 1.5 that gave SOTA results on various reasoning benchmarks, widely surpassing GPT-4 and other existing models.
• RLHF: Part of the core-modeling team that builds post-training recipes integrating human labelled data into reward models and performing RLHF to improve the model across all pillars like quality, factuality, coding, reasoning, instruction-following etc.
• User Signals: Developed data collection and training recipes to extract useful signals from production user’s activity on Gemini, and use the extracted user-signals to align Gemini models towards produc- tion user preference.
• Trained multiple Gemini models that served production users.

Nov 2020 – Apr 2023 · 2 yrs 5 mos

• Google Assistant
• Built and launched high quality latency-sensitive ML models to power conversational queries
• in Google Assistant, making assistant more friendlier and human-like.
• Designed improved distillation and data generation recipes to distill LLMs like Gemini, MUM and LaMDA into smaller models.
• Reduced end-to-end Assistant query latency by 110ms on iOS and ChromeOS by simplifying the existing infrastructure.

Oct 2018 – Nov 2020 · 2 yrs 1 mo

• Wear OS
• Led and owned the design, development and testing of 1P-Experiences like Proactive Assistant and Google Maps on the Google Pixel watch and other WearOS devices.
• Presented the work at Google’IO 2019. The Google Maps app also received praises in various media articles.

Jun 2017 – Jul 2018 · 1 yr 1 mo · Hyderabad Area, India

• Azure Site Recovery
• Optimized download speed of HRL log files by a speedup of 500% through an
• efficient producer consumer download mechanism, thereby reducing total
• replication time.
• Designed and implemented Blob Reader, IO Throttler and Circuit Breaker
• modules for HRL Processor.
• Improved overall time in A2A replication scenarios by analyzing and improving
• timeout and retry strategies and buffer windows.

May 2016 – Jun 2016 · 1 mo · Hyderabad, India

• Designed a framework to allow Azure IaaS VMs to take app-consistent backups for various workloads running in the backgroud. The framework allowed user run a pre / post script for the respective application before and after the VM Snapshot to allow the user to take app-aware and app-consistent Azure IaaS VM Backups. The script takes advantage of native app API's to put the app in a consistent state. The framework was also implemented and tested for Linux IasS VMs.

Jul 2015 – May 2017 · 1 yr 10 mos · Roorkee

• Successfully managed and organized INSOMNIA - an international level coding competition and promoted coding in campus and outside by organizing various lectures and programming contests.

Jan 2014 – Jan 2015 · 1 yr · Roorkee

• Worked on Campus Omegle - a campus clone of Omegle, and Erdos - a math challenge site for campus contests.