Hi, I'm Pratik — a Research Engineer at Meta in New York, working on foundation models for Instagram's recommendation systems. My work spans large-scale sequential modeling, multimodal architectures, and real-time ranking systems serving billions of users.

Before Meta, I completed my Master's at UC Berkeley (MIMS, ML focus), interned at Amazon Science on graph-based recommendation systems, and spent three years as an ML/Software Engineer at JPMorgan Chase and Tingtun AS (Norway).

I'm interested in building intelligent systems that operate at scale — where model quality, efficiency, and real-world impact are all first-class concerns.

Experience

Nov 2023 — Present

Research Engineer, ML · Meta

New York, NY

  • Worked on HSTU, Meta's LLM-like self-attention based sequential retrieval foundation model for Instagram Reels. Designed a Sequential Interest Encoder via cross-attention — improved offline share Recall@50 by XX% and drove a XX% increase in Reels sharing post-launch.
  • Leading design of HSTU for Instagram Feed retrieval, integrating hierarchical mixed gating and multimedia-type ranking across photo, video, and carousel — resulting in a XX% increase in time spent on Instagram.
  • Delivered efficiency improvements of approximately XX through training data sampling, sequence length reduction, and dimensionality reduction.
PyTorch Foundation Models Transformers Sequential Modeling Recommendation Systems
May 2022 — Aug 2022

Applied Scientist Intern · Amazon Science (Alexa)

Remote

  • Worked on graph-based machine learning recommendation systems for the Alexa team, focusing on personalized content discovery.
Graph Neural Networks Recommendation Systems Python AWS
Jul 2018 — Apr 2021

Machine Learning Engineer · JPMorgan Chase & Co.

Mumbai, India

  • Developed and deployed aspect-oriented sentiment analysis and customer satisfaction forecasting on AWS SageMaker, guiding product actions for Chase Bank credit cards.
  • Built customer churn prediction models using ensemble methods with XX% accuracy, enabling targeted retention campaigns.
  • Built Java-based data pipelines processing 1M+ daily transactions for accurate, timely generation of margin calls.
NLP ML Engineering AWS SageMaker Python Java

Projects

GNNIE: GNN Recommender-as-a-Service

Built a scalable graph neural network based recommendation system for personalized and interpretable recommendations. Awarded the James Chen Award at UC Berkeley, 2023.

  • Graph Neural Networks
  • PyTorch
  • Recommendation Systems
  • Python

Benchmarking DRL for Hyperspectral Band Selection

Evaluated deep reinforcement learning approaches for selecting informative spectral bands from high-dimensional hyperspectral imagery in an unsupervised setting.

  • Reinforcement Learning
  • Computer Vision
  • Python

Publications

IEEE · ICCUBEA 2018 2018

Automated Aspect Extraction and Aspect-Oriented Sentiment Analysis on Hotel Review Datasets

Used NLP techniques to extract aspect-based ratings for hotel reviews and perform aspect-oriented sentiment analysis to summarize reviews into a structured rating system.

NLP Sentiment Analysis Aspect Extraction
Turkish Journal of Computer & Mathematics Education 2021

Need-Based Access Control Framework for an Emergency Response System

Proposed and evaluated an ML-based need-driven access control model for emergency response scenarios, balancing security and situational urgency.

Access Control ML Security
Innovation Patent · IP Australia 2021

A System and Method for Need-Based Access Control Framework for an Emergency Response System

Innovation patent granted by IP Australia for an ML-based access control system designed for emergency response environments.

Patent ML Access Control