As a machine learning lead at Amazon Music, I led advanced ML initiatives to improve search and retrieval. Proven success in developing high-impact solutions such as low-latency deep learning model inference service, generative transformer model based spell correction etc. with multiple industry-recognized publications and significant metric improvements.

As a Senior Machine Learning Engineer at Meta, I led cross-team efforts to enhance FB Marketplace’s personalized retrieval, collaborating on feature engineering and ranking model optimization. I also designed a robust ML stack for content integrity, deploying few-shot classification and multimodal vector search to detect misinformation on FB and IG.

As a Staff Data Scientist at Walmart Labs, I drove the vision for search query understanding and ranking models, building foundational search systems from the ground up for Walmart Canada. I collaborated closely with product leadership to prioritize impactful machine learning initiatives and played a key role in hiring and mentoring ML talent across multiple verticals, showcasing leadership in both technical and strategic capacities.

At Rakuten Inc., I helped developed a scalable short-text classification service for product taxonomy mapping and collaborated in hiring and onboarding a new machine learning team.

Click Prediction System: Researched, developed and build end to end data pipe line for click prediction system for video ad campaigns that classifies high performing bid request segments from different ad exchanges. The system involves setting up a large scale ad log processing Spark based infrastructure for generating training data. The model is trained using Vowpal Wabbit online learning framework exploiting stochastic gradient descent, adaptive learning, per feature learning rate, L1 regularization for sparsity, Online Importance Weight Aware Updates, non linear features. The features weights are uploaded in Redis to be consumed by RTB Bidder. Technologies used: IPython, Scipy Stack, Scala, C++, Spark, Redis, Vowpal WabbitSpark Reporting Pipeline: A stand alone spark cluster was set up to read bid, impressions and events logs from S3/hdfs and process them using Spark based jobs. The Spark job generates different customer segments and computes various reporting statistics like clicks occurred, view completion rate, impressions served, win rate etc for each segment.

C1X is a programmatic direct technology company, it is a direct guaranteed buying and selling of online advertisements platform aiming towards automation and efficiency.RTB Pacing and Optimal Bid Price Algorithms • The smooth pacing algorithm probabilistically ranks campaigns given an ad request in real time environment. The aim of the algorithm is to prioritize campaigns according to active budget, targeting constraints and available inventory. • Developed a Thompson Sampling based Dynamic Bayesian Bandit algorithm that can efficiently find the optimal bid price of a segment with desired win rate. The algorithm continuously explores the bid price landscape to see if the RTB environment has changed i.e. if the win rate distribution for a particular cpm bid price has changed or not and readily adapts to the change using beta posterior variance inflation based technique. Ad Inventory Forecasting • Built a time series forecasting algorithm for ad inventory prediction for guaranteed ad campaigns. • Research included forecasting algorithms like Holt-Winters algorithm, ARMA, Fourier Series and Dynamic Linear Models. • Python based development of algorithm and Java based deployment of REST service.Ad Log Processing • Set up SPARK based infrastructure and daily ad and DMP log processing jobs.

My most recent work at Tremor was concerned with development of a revenue optimization system for distributing ads to different publishers with constraints on ad performance, inventory, dispersion and target revenue. This is a variant of multi knapsack optimization algorithm. Initial experimentation involved use of Linear Programming techniques but the final model used a variant of Markov Chain Monte Carlo sampling technique to distribute ads to various publishers.

My first project at GumGum involved keyword extraction from a web page for ad targeting . Keyword extraction involves extraction of ngrams and entities from wikipedia corpus and storing idf scores in memcache. The keyword extraction service parses the web page for unigrams,bigrams, trigrams in body as well as header and title, it uses a scoring model based on tfidf linear weighted model to extract important keywords.The second project involved development of anomaly detection service, that given a page url, will try to detect whether the webpage is safe for targeting or not. If unsafe, the classifier will try to predict the class of the unsafe document e.g. is the page containing news related to death, crime etc. Technologies used : libsvm, weka, Java, Groovy

Authentication system security for Pega product on Cloud: My Current work at Pega, comprises of building a secure authentication system for Pega product on cloud. We have designed and implemented an exponential delay based system to counter brute force attacks, captcha based malware protection and web service security techniques, securing the cloud authentication from active attacks. Cache Optimization: As a part of Performance Optimization team at Pega, my current ongoing sprints are concerned with improving the performance of different cache’s initialized during Pega Server startup. PRPC runs on clustered environment where each node operated independently but are connected to a single database. During start up each node initializes a number of caches like Meta data related cache, Assembly cache etc. My current work involves with optimizing the boot time of the server by making cache initialization as fast as possible.

Financial Reporting Technologies August ‘11- Arpil ‘12 My work at BOA Continuum Solutions was mainly concerned with supporting and developing Financial Reporting technologies. I participated in the development of Corporate Reporting. It’s a web hosted application that helps in maintaining transaction details (a formal ledger) for equity, derivatives and fixed income investments by various specialist. It’s a java based application written primarily in Applets, using MS Excel and Oracle 10g as database.