Oct 2024 – May 2025 · 7 mos · Montreal, Quebec, Canada · On-site

• At ServiceNow Research, I led research on RETREEVER—a tree-based, coarse-to-fine retrieval framework that learns a differentiable binary hierarchy over text embeddings and optimizes routing end-to-end for retrieval, and am the first author on the research manuscript describing this approach. The work shows RETREEVER preserves fine-level accuracy, offers stronger coarse representations for low-latency search, and yields an inspectable hierarchy with meaningful semantic groupings, achieving favorable accuracy–latency trade-offs among hierarchical methods.

May 2024 – Oct 2024 · 5 mos · Montreal, Quebec, Canada · On-site

• My research on asynchronous (irregular) time series led to an ICML 2025 publication—“LAST SToP: Last Stop for Modeling Asynchronous Time Series”—with me as first author. I developed a simple, scalable framework for asynchronous (irregular) time series and a coarse-to-fine extension of prompt tuning to adapt LLMs to this setting. The approach unifies forecasting, imputation, and anomaly detection in one recipe, leverages pretrained LLM prior knowledge, and remains efficient. As first author of the ICML 2025 paper “LAST SToP: Last Stop for Modeling Asynchronous Time Series,” I designed the core algorithms, training objectives, and evaluation protocol. Results show strong generalization across multiple datasets with balanced accuracy–latency trade-offs, enabling use in low-resource environments.

Jun 2021 – Jul 2023 · 2 yrs 1 mo · San Francisco Bay Area · Remote

• As a member of the Overdub Team, I am responsible for researching, implementing, and deploying cutting-edge deep learning algorithms for speech editing, text-to-speech (TTS), and voice cloning. My work in generative machine learning involves utilizing a variety of modeling techniques, including variational autoencoders (VAEs), generative adversarial networks (GANs), diffusion models and LM based audio autoregressive generation models.

Jan 2020 – Jun 2021 · 1 yr 5 mos · San Francisco Bay Area

• As a member of the Merchant Fraud Team, I contributed to the design and improvement of models to detect fraud among new merchant sign-ups. One of my notable achievements was leading a working group that combined text embeddings from models like Fasttext, BERT and its variants, with traditional NLP techniques like count and tf-idf vectorizers to add text-based features to our merchant fraud classification models.
• Furthermore, I improved our existing Anomaly Detection framework for catching fraud attacks, resulting in a 72% reduction in MTTD (mean time to detect) while increasing the number of attacks caught by 45%. I also identified and patched blind spots in our model scoring pipeline, resulting in faster human intervention on merchants and preventing payouts of 39% more fraudulent balance at a similar precision. Overall, my contributions have helped to significantly improve the effectiveness of our fraud detection systems.

May 2018 – Dec 2019 · 1 yr 7 mos · Greater Los Angeles Area

• As a Research scientist at Retention Science, I was responsible for designing, prototyping, and deploying generalized machine learning models that empowered over 145 e-commerce companies to execute successful email marketing campaigns.
• I led a team that was accountable for building and maintaining more than 30 predictive machine learning models, including lead scoring, churn prediction, lifetime value modeling, recommendation systems, and real-time multi-armed bandit testing.
• Our models powered over 4 billion daily predictions for 385 million users across various industries, such as fashion, food, and books, among others, and helped our clients achieve significant growth in revenue and customer retention.

Sep 2014 – Mar 2018 · 3 yrs 6 mos

• At Snapchat, I've had the opportunity to work on several impactful projects, including:
• Smart Content Precaching: As a lead of a team of 3, I contributed to designing, developing, and extensively AB testing machine learning and rule-based strategies for deciding which content to preload for a user based on both global trends and personalized behavior. My contributions resulted in a significant improvement in the utilization of precached content (from 10% to 25% and 60% to 80% for different content types), improved content engagement by 5%, and reduced loading screens by 50%. In addition, I helped to grow content MAU by 7% and reduced user churn in countries like SA from 15% to below 5%.
• Friend Feed Ranking: I played a key role in designing, developing, and extensively AB testing models for friend feed ranking, as well as implementing efficient cloud dataflow pipelines to produce ranking signals for over 150M users. By analyzing user data to find relevant signals and developing heuristics based on those signals, I was able to increase user engagement on stories by 5% snap views.
• Snapchat Android: In this role, I had the opportunity to implement various features for the Snapchat Android app, including designing and implementing the serving stack for content precaching and implementing the Android component. Additionally, I implemented the ad server to support early monetization strategies adopted by Snapchat and various product features for Discover and Stories. I also designed and implemented the serving stack for subscriptions to Discover editions and implemented the Android component.

Oct 2011 – Sep 2014 · 2 yrs 11 mos

• At Google, I worked on several projects spanning across different domains, some of which are highlighted below:
• Knowledge Graph(KG) (http://goo.gl/rNjvGP) aimed to remember structured information about entities known to Google. As part of the Recon(reconciliation) project, I designed and implemented a classifier to identify entities in the data source as new or not new, preventing duplicates in KG. This reduced losses in candidate prediction by 50%.
• Wikisets performed information extraction from lists and tables in Wikipedia and asserted them in KG. I designed and implemented the extraction of entries not linked on Wikipedia (black links), resulting in 18% more extractions. I also implemented heuristics that increased recall of the pipeline by 10 times at threshold precision and an additional regression to score extractions, increasing recall two times at threshold precision.
• In the Mobile Tools domain, I enabled text-level diffs for Google search in low-end mobile and desktop browsers to catch development errors. I also designed and implemented an edit-distance based algorithm for finding differences in Google search screenshots to catch visual errors in Google search.
• As part of the Mobile Transcoder project, which is a web service that allows users of low-end mobile phones to access desktop web pages in a friendly format, I designed and implemented a text processor to specifically handle text files. This resulted in a reduction of transcoder traffic losses by about 12%.