Nov 2024 – Mar 2025 · 4 mos · Seattle, Washington, United States · Remote

Feb 2024 – Oct 2024 · 8 mos

• Leading features team for Ads recommendation whose responsibilities include the following:
• ◆ Manage the design and optimization of data pipelines for collecting, preprocessing, and transforming large volumes of data
• ◆ Develop real-time monitoring and alerting systems to proactively identify and resolve issues. Define KPIs and develop dashboards to track system performance and highlight areas of improvement.
• ◆ Conduct exploratory data analysis to identify patterns and insights that can enhance model performance
• ◆ Oversee feature engineering and selection to extract meaningful information from raw data

Apr 2022 – Jan 2024 · 1 yr 9 mos

• 𝐂𝐨𝐧𝐭𝐞𝐧𝐭 𝐋𝐢𝐟𝐞𝐜𝐲𝐜𝐥𝐞: This involves touching upon different stages a content goes through viz. Cold start, growth, maturity and expiration. The efforts ensure a healthy system that facilitates content growth and in turn enhances user and creator satisfaction. Led to following impact:
• 📌 Improved the performance of cold start recommendation by~30%
• 📌 Increased the usable corpus size by 10X
• 📌 Proposed the idea of reusing supply to reduce expenditure on content sourcing and bootstrapping seasonal events
• 📌 Development and utilisation of HAMSA - a highly scalable vector similarity search framework, for business use cases & effective targeting of content on the platform
• 𝐂𝐨𝐧𝐭𝐞𝐧𝐭 𝐌𝐚𝐫𝐤𝐞𝐭𝐩𝐥𝐚𝐜𝐞 𝐎𝐩𝐭𝐢𝐦𝐢𝐬𝐚𝐭𝐢𝐨𝐧: Multi-objective ranking/recommendation techniques, quantifying stakeholder objectives, developing user understanding modules and developing joint optimization modules

Nov 2020 – Mar 2022 · 1 yr 4 mos

• 𝐂𝐚𝐭𝐞𝐠𝐨𝐫𝐲 𝐆𝐫𝐨𝐰𝐭𝐡: Led a team of SDEs, MLEs and worked with the product team to lay out the growth plan for identified categories on ShareChat. Some of the categories we worked on are News (+Hyperlocal News), Cricket and long-form-videos. Few things my team did as a part of this are mentioned below:
• ★ Bootstrapping the content category
• ★ Generate personalised recommendations
• ★ AB tests to identify the marginal gains associated with the category
• ★ Create roadmap for discoverability and expansion
• 𝐏𝐨𝐬𝐭 𝐋𝐢𝐟𝐞𝐜𝐲𝐜𝐥𝐞 & 𝐂𝐫𝐞𝐚𝐭𝐨𝐫 𝐒𝐮𝐜𝐜𝐞𝐬𝐬: Create an unbiased system to facilitate healthy growth of both content and creators across the platform.

Jan 2019 – Oct 2020 · 1 yr 9 mos

• ◆ 𝐃𝐮𝐩𝐥𝐢𝐜𝐚𝐭𝐞 𝐃𝐞𝐭𝐞𝐜𝐭𝐢𝐨𝐧: Duplicates harm the system in multiple ways which has a direct impact on the health & growth of the platform. Developed a scalable end-to-end system for identifying duplicates in a corpus of 50M+ content pieces. Explored and leveraged fast, scalable SOTA nearest-neighbour search implementations.
• ◆ Led multiple threads as a part of one of the biggest 𝐜𝐥𝐨𝐮𝐝-𝐭𝐨-𝐜𝐥𝐨𝐮𝐝 𝐦𝐢𝐠𝐫𝐚𝐭𝐢𝐨𝐧𝐬
• ◆ Developed an end-to-end OCR pipeline that acts as the source for multiple tasks like fake news detection, hate speech detection and enhancing recommendation by enriching metadata with text features.

Sep 2017 – Dec 2018 · 1 yr 3 mos

• ✪ 𝐂𝐨𝐧𝐭𝐞𝐧𝐭 𝐌𝐨𝐝𝐞𝐫𝐚𝐭𝐢𝐨𝐧: Best-in-class multimodal AI algorithms for automated moderation of multimodal content with minimal response time. We have shared the details in the multi-part blog series.
• ✪ 𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫𝐢𝐧𝐠 𝐓𝐫𝐞𝐧𝐝𝐢𝐧𝐠 𝐓𝐨𝐩𝐢𝐜𝐬: Developed an end-to-end module for identifying and serving most trending topics across the application based on user interaction data.

Jul 2019 – Jan 2021 · 1 yr 6 mos

Jun 2016 – Aug 2017 · 1 yr 2 mos · Bangaon Area, India

• » 𝐂𝐮𝐬𝐭𝐨𝐦𝐞𝐫 𝐈𝐧𝐭𝐞𝐧𝐭 𝐌𝐨𝐝𝐞𝐥𝐢𝐧𝐠: Classify the intent of a customer into one of the pre-defined intentions as early as possible during the session for a better experience. The end goal is to save resources wasted in misdirected calls.
• » 𝐒𝐦𝐚𝐫𝐭 𝐒𝐞𝐚𝐫𝐜𝐡: Understand user query and map effectively to the potentially correct destination to cut short the process of browsing and searching.

May 2015 – Jun 2015 · 1 mo · Bengaluru, Karnataka, India

• Count Regression Models and Variable Selection
• ◘ Built a count regression model which can be used for prediction in cases when the response variable takes only non-negative values.
• ◘ Explored different variable selection techniques which can be used for modelling count data.
• ◘ Analysed the results of various selection techniques and combined the results to obtain a final list of variables from which variables can be obtained for modelling.
• ◘ Explored different count regression models like poisson regression model, quasi-poisson regression model, negative binomial model, zero-inflated model and hurdle model.
• ◘ Compared the results of different count models with the logistic regression model.
• ◘ Showed what extra information count regression models are able to add as compared to the logistic regression model.
• ◘ Finally concluded that count regression models capture some vital information that logistic regression model is incapable of capturing.

May 2014 – Jul 2014 · 2 mos · Klagenfurt,Austria

• Sequential Estimation Procedures For Burn-In Testing
• ☛ Studied and learnt the use of Bayesian methods in sequential testing of semiconductor products, which are vulnerable in their early life, by means of a burn-in (BI) study.
• ☛ Used the data of a case study to monitor a clinical trial and compared the results obtained using carefully elicited priors with those obtained using vague priors.
• ☛ Used the backward induction technique to arrive at a Bayesian optimal decision at a time point other than the last one.
• ☛ Automated the above mentioned technique through a R script. It plots a graph between posterior expected loss and the prior and helps in arriving at an optimal decision by giving a continuation region.

Jun 2013 – Jul 2013 · 1 mo · Mandi Area, India

• ★ Explored different options trading strategies which can be used in different market conditions to limit risk and maximize the return.
• ★ Observed the change in current stock price, the strike price, interest rate, volatility and time of maturity using modern day techniques such as the Black-Scholes-Merton model.
• ★ Explored the Greeks which are the Delta, Rho, Vega, Theta and Gamma to study the sensitivity of the option prices to the various input parameters.
• ★ Applied Markowitz Portfolio model in R statistical software to find an optimal portfolio from combinations of portfolios by plotting the efficient frontier using historical data.