Apr 2024 – Jun 2025 · 1 yr 2 mos · Bengaluru, Karnataka, India · On-site

Feb 2020 – Dec 2022 · 2 yrs 10 mos

• Product Trust Score (PTS) to enhance Selection and Ranking in Amazon Choice, Recommendation and Search.
• Launched a Unified Global Deep Learning (DL) model in 16 countries world-wide; measured annualized business impact of ~50 MM $ in an online A/B test, after integration in just 4 out of 16 countries.
• Designed a principled and data driven approach for identifying “what to share?” in Multi-Task
• Learning setting using Fisher Information in weight parameters.
• Designed a retraining pipeline based on Continual Learning (CL) approaches to adapt to
• local country level changes while maintaining performance on other countries.
• Mentored a team of scientists and interns to improve on Cold Start products using Graph Convolution Networks and advanced dropout strategies and design a novel scalable Uncertainty Estimation method based on quantile regression.
• Responsible for yearly planning for PTS; includes identification of problem statements for
• offline experimentation and research.
• Received accolades from team members for effective guidance and skill sharing, and from
• manager for general effective management of the project and global launch.

Feb 2018 – Feb 2020 · 2 yrs

• Influencing Bidding decisions via. user engagement prediction
• Leveraged recent user engagement across webpages to infer future engagement on a specific
• webpage. User engagement probabilities were used for improving real time bidding decisions.
• Improved engagement models on cold URLs by leveraging URL hierarchies and implementing
• hourly aggregation pipelines to get more recent URL level engagement information.
• Explored Semi-Supervised methods to handle partially labelled data and latent variable
• survival methods to predict duration of stay and hence improve viewability models.
• Provided guidance to scientists and engineers in the team for building models and pipelines
• to predict various user engagement signals.
• Collaborated cross-functionally to integrate user engagement metrics in Amazon Ad offerings.
• Various model improvements lead to 10% and 33% increase in ‘High Engagement’ ad inventory
• for display and video ads respectively.
• Detecting robotic or malicious activity on Amazon Ads
• Collaborated with other scientists in designing a Collusion Detection algorithm by applying
• Tensor Factorization and Sequence modelling approaches on customer Ad interaction data.
• Improved label quality using weak supervision, label propagation, etc, for a back bone supervised classification model acting at the ad click level. Enabled use of high cardinality
• categorical features into the model which otherwise would have biased the model.
• Designed a counter-factual based root causing algorithm which recursively breaks a macro
• anomaly into smaller anomalies and hence finds a small set of problematic entities.
• Collusion Detection and Counter-Factual based Anomaly Detection algorithms together led to
• ~20MM $ of saving in bid value for the Advertisers annually.

Jul 2017 – Feb 2018 · 7 mos · Bangalore

• Productionized a CTR and CVR model for company’s DSP product; normalized for position
• bias; used URL hierarchy to deal with cold start URLs; explored Deep Factorization Machine
• to learn from very sparse data.
• Explored topic modelling and word embedding approaches to learn URL text representations
• used in CTR/CVR models of a SSP product.

Apr 2016 – Jul 2017 · 1 yr 3 mos · Mumbai Area, India

• Created a new user recommendation engine leveraging user demographics; led to 50%
• increase in conversion and 20% increase in retention of new users.
• Developed Customer-Match propensity model; and collaborated with marketing head to
• devise a deals targeting strategy based on the score; 2% increase in net entry fee collected.
• Designed a Collusion Detection algorithm to mitigate referral fraud; identified suspect
• clusters offline and blocking same cluster joins; saved INR 1M in cash bonus.
• Collaborated on SOP docs for setting possible fraudulent behaviours that manual annotators
• should look for in flagged accounts and hence generate labels for supervised model.
• Took sessions on various topics in Machine Learning for other scientists; Linear Algebra,
• Statistics and Probability, Machine Learning Basics, Graphical Models, etc.

Jul 2015 – Apr 2016 · 9 mos · Mumbai Area, India

• Developed an Item-Item association model for Cross-Selling; 10% increase in orders
• Devised Churn prediction model to priorities delivery; 20% increase in long distance orders.
• Techniques used: Gradient Boosted Model, K-mean Clustering, User-User Collaborative filtering, Low-rank Matrix Factorization by Latent feature learning, Item-Item Association

Jul 2012 – Jun 2015 · 2 yrs 11 mos · gurgaon

• Responsible for creating Collection models i.e. probability of payback by a defaulted customer. Calls or letters were sent to the customers based on the model output.
• Employed topic model to extract early financial stress indicators from historical call records. 9% increase in payer capture rate; received Analyst of the Quarter Award.
• Later on also responsible for Time-Series Forecasting of Account Receivables and hence use them for calculating reserves required by American Express in adverse market scenarios.
• Designed an approach assuming Markovian Transition of Account Receivable from one state to another to forecast losses in future and hence decide about the reserves required.
• Developed a Sensitivity Analysis method to determine importance of each market index on the forecasts and hence the reserves required.
• Recognized by Chief Risk Officer at American Express for delivering next generation of Capital Adequacy Models.