Feed Recommendations Team

• Developing e2e improvements across the recommendation funnel to improve user interest understanding and candidate relevance, applying techniques such as collaborative filtering, clustering, and epsilon-greedy exploration.

• Data-driven analysis on metrics such as user coverage, engagement CTR, content freshness, topic diversity.

Reality Labs - Cognitive AI team

• Developed model-free solutions based on Transformer models (BART and Decision Transformer) designed to solve

problems relating to sequential decision making (SDM) and planning for applications in the AR space.

• Designed and developed a flexible and generic data serialization pipeline using Pytorch Text that is extendable to different domains and model-free architectures. The pipeline handles the end-to-end serialization of raw metadata into the input format required by downstream models for training and inference.

• Trained and benchmarked the developed model against seven baseline agents, leading iterative improvements which resulted in the model achieving the highest session return, demonstrating its effectiveness in sequential decision-making tasks.

Meta Superintelligence Labs - Metamate AI Assistant ML

• Owned end-to-end development of an internal AI assistant’s response quality evaluation system using LLM-based methods, from metric design through productionization, enabling objective quality tracking across 100% of response traffic with precision/recall exceeding 70%.

• Developed automated benchmarks for evaluating tool-call orchestration across diverse use cases, i.e. coding, search, writing, artifact editing.

• Initiated and facilitated team knowledge-sharing initiatives including ML paper reading groups and mentorship programs, accelerating onboarding for new team members and driving adoption of state-of-the-art techniques.

• Led and published research examining how domain-specific terminology influences the performance of Transformer-based meeting summarization models. Curated a set of gold-standard annotations for domain terminology within the ICSI meeting transcript dataset (75 transcripts) to support further research and model development.

• Developed and evaluated sliding window methods to generate meeting minutes automatically using BART-based abstractive summarization, which eliminates the need for annotated training labels.

• Led the coordination of efforts and progress within the virtual team. Developed the overall data processing and evaluation pipeline using PyTorch and enabled data visualization using R.

Developed features for Jobs on Facebook using technologies such as React, Hack/PHP, GraphQL.

Developed features in Java Server Pages for an integrated eligibility web app for users to determine their eligibility towards various benefits.