How do we know an agent is ready for real customers? I build evaluation systems that connect offline simulation, human judgment, and online experiments.
A unified framework for context engineering, prompt optimization, offline evaluation, and production validation across five Nubank deployments.
+37 percentage points in transactional NPSCard-delivery A/B test, compared with prior agent variants.
How do reward design and evaluation shape model behavior after pretraining? I study where standard preference-optimization recipes become brittle and how to make alignment more controllable.
Introduces an alpha-parameterized reward shape that controls likelihood displacement and over-optimization, improving alignment performance over SimPO and DPO across Mistral-7B and Llama-3-8B.
How do we preserve model quality while meeting real training, latency, and serving budgets? I work across distillation, compression, quantization, and systems-aware inference.
Distills and compresses production language models with structured pruning and quantization, retaining much of larger-model quality while reducing training and serving cost and latency across recommendation workloads.
How do we optimize recommendations against global business, allocation, and diversity constraints at web scale? I develop optimization methods that turn model predictions into globally coordinated decisions.
Presents DuaLip, a distributed solver for linear programs with trillions of variables, deployed across LinkedIn recommender systems and released as an extensible open-source library.
CIKM 2026, half-day tutorial Nov 2026
The Web Conference (WWW) 2025, half-day tutorial, Sydney, NSW, Australia May 2025
AI Engineer, with Shreya Rajpal Jul 2026
video1.9K views
QCon AI New York Dec 2025
NuSession Silicon Valley Jun 2026
ICLR SPOT workshop Apr 2026
Lead company-wide research on AI agents, foundation models, LLMs, and large-scale constrained optimization for more than 135 million customers.
Tech lead for compressing and distilling LinkedIn's ranking foundation model, and manager of the AI foundations optimization team working across Feed, Ads, Jobs, Notifications, and Growth.
Deep learning for computer vision and autonomous systems.
Scalable services for infrastructure automation and security on the E-Commerce Platform Services team.
Research at the Language Technologies Institute, supported by a Graduate Research Fellowship.
AIMA All India Merit Scholarship, All India Rank 2, and a BITS Pilani merit scholarship.
An incomplete list of people whose guidance shaped how I think about research. I owe each of them more than a line.