Snacks Weekly on Data Science

In this episode, we’ll explore RazorPay, its business, and the critical role customer service plays in its success. We'll dive into how RazorPay revolutionized customer ticket categorization using generative AI. By replacing customer-selected categories with an AI-driven system, they enabled automatic interpretation of ticket details. This approach incorporated pre-processing, prompt engineeri...

In this episode, we will explore the foundational concepts of causal analysis, focusing on its two main pillars: causal discovery and causal inference. We will discuss the types of questions these pillars aim to answer and provide illustrations of related methodologies to better clarify their concepts. For more details, you can refer to their published tech blog, linked here for your reference:  h...

In this episode, we will explore how Noom developed a customized AI assistant solution using generative AI models. By incorporating key components such as prompt engineering, dynamic personalization, and integration with their knowledge base, the team created a smarter, more reliable assistant to support their customers. This approach offers valuable insights for anyone looking to build tailored g...

In this episode, we will explore Airbnb’s approach to personal data classification. We will begin by introducing the importance of protecting personal data. Next, we will examine the technical solution, where the three pillars—Catalog, Detection, and Reconciliation—form the backbone of their workflow. Finally, we will discuss how performance metrics are implemented to ensure the system remains rel...

In this New Year’s episode, I reflect on one of my earliest data science projects. My son will “interview” me about my experiences during that time. The key project discussed in this episode is also featured in a LinkedIn article I wrote a few years ago (⁠ https://www.linkedin.com/pulse/my-first-data-science-project-pan-wu/⁠⁠ ). I’d love for you to check it out, leave a 5-star review, and subscrib...

In this episode, we will explore Netflix’s approach to content recommendation using contextual bandits and reward engineering. We will also discuss the important role of proxy reward functions and how Netflix leverages offline machine learning models to predict delayed customer feedback, enabling them to continuously improve their recommendation engine and deliver a more personalized viewing exper...

In this episode, we will introduce the concept of paid marketing. We’ll explore how Instacart’s team developed an adaptive experimentation framework that continuously balances exploration and exploitation, ultimately maximizing marketing efficiency. For more details, you can refer to their published tech blog, linked here for your reference:  https://tech.instacart.com/bandits-for-marketing-optimi...

In this episode, we will introduce what CVS Health is and the importance of product recommendations for their business needs. We will delve into how their data science team leveraged advanced technologies, including Graph Neural Networks and generative AI models like GPT-4, to develop a prototype system to make product bundle recommendations. For more details, you can refer to their published tech...

In this episode, we will introduce the issue of data imbalance and its impact on machine learning models, especially for text data. We will discuss a range of augmentation techniques and walked through how Walmart’s data science team built an automated augmentation module to apply these techniques consistently and effectively. For more details, you can refer to their published tech blog, linked he...

In this episode, we will discuss what is Instamart, its business model, and the need to optimize operational efficiency continuously. We will explore how the team tackled picker assignment issues, tested a multi-order batching solution, and applied modeling techniques to enhance Instamart’s speed and service quality. For more details, you can refer to their published tech blog, linked here for you...

In this episode, we introduce GovTech Singapore and its reasons for tackling the content moderation problem. We discuss their innovative approach to building the moderation classifier, which involves using a consensus voting mechanism with existing commercial LLMs to improve labeling in the training dataset, providing a strong foundation for developing the machine learning model. For more details,...

In this episode, we will introduce Afresh, explore the concept of demand forecasting, and discuss the critical role of promotions in grocery forecasting. We will also examine Afresh’s technical solution, which combines deep learning with a carefully curated set of data inputs to create an accurate, promotion-aware demand forecasting system. For more details, you can refer to their published tech b...

In this episode, we will explore the business model of Booking.com and the unique challenges it faces in preventing fraud. We will discuss how graph technology can enhance fraud detection by representing data through relationships and how combining this with machine learning enables Booking.com to detect and stop fraudulent activity in real-time. For more details, you can refer to their published...

In this episode, we will introduce a key marketing challenge—attribution—and explore how marketing mix modeling (MMM) can help solve it. We will also discuss how MMM works in practice and examined the trade-offs between using consultants for MMM development versus building an in-house solution. For more details, you can refer to their published tech blog, linked here for your reference:  https://m...

In this episode, we will explore Urban Company’s business needs and the role of causal machine learning in addressing them. We will delve into three key models—S-learner, T-learner, and X-learner—and used a simple example to illustrate how each one works. These models provide valuable insights by offering more accurate estimates of cause-and-effect relationships, helping companies make better data...

In this episode, we will explore how Faire tackled the challenge of product categorization. They initially used the K-nearest neighbor algorithm with CLIP embeddings, which improved categorization but still required manual corrections. To further enhance accuracy, the team fine-tuned a vision-language model using their in-house dataset, increasing accuracy significantly. This solution showcases ho...

In this episode, we will discuss why Walmart relies on online experimentation to drive data-driven decisions and how reducing variance is a crucial challenge to making these experiments more efficient. We will also introduce the CUPED methodology, explaining how Walmart leverages it to speed up its experimentation process, enabling faster, more reliable insights for continuous improvement. For mor...

In this episode, we will explore the challenge Netflix faces in building machine learning models for video understanding. We will examine Netflix’s solution—a self-service system with active learning that empowers video experts to participate in creating and refining machine learning classifiers through a streamlined, three-step process. For more details, you can refer to their published tech blog...

In this episode, we explore the concept of incrementality in marketing, and how the data science team at Expedia leverages geo-testing to successfully measure their marketing campaign’s  incrementality  For more details, you can refer to their published tech blog, linked here for your reference: https://medium.com/expedia-group-tech/measuring-marketing-success-the-power-of-incrementality-and-geo-t...

In this episode, we explore the concept of out-of-app marketing and its importance for businesses like Uber. We'll discuss the challenges involved in personalizing out-of-app marketing messages, the essential components of their recommendation architecture, and how the team created customized solutions to improve personalization and relevance in their recommendations. For more details, you can...

In this episode, we will discuss the importance of ETA for ridesharing apps and the challenges of providing a reliable ETA to users upfront. We delved into the practices of the machine learning team at Lyft, examining how they developed a solution to address this unique challenge using a lightweight machine learning model. For more details, you can refer to their published tech blog, linked here f...

In this episode, we will explore how Yelp navigates the challenges of incorporating video reviews into its platform. We will discuss the use of machine learning to detect inappropriate content and the strategies to maintain the quality and integrity of its platform. For more details, you can refer to their published tech blog, linked here for your reference: https://engineeringblog.yelp.com/2024/0...

In this episode, we will introduce what a brand is and why measuring brand perception is crucial. We will discuss how social media data can be used to achieve this and look at how Airbnb leverages deep learning methods and word embedding technologies to quantify brand perception more effectively. For more details, you can refer to their published tech blog, linked here for your reference: https://...

In this episode, we will explore the concept of quasi-experimentation and its role in Instacart's decision-making process. We will take a closer look at one specific methodology, regression discontinuity design, explaining its key concepts and demonstrating how it works through an interesting example. For more details, you can refer to their published tech blog, linked here for your reference:...

In this episode, we will discuss the machine learning architecture built by LinkedIn for their premium product recommendation. We will explore the machine learning architecture, which includes a two-towered neural network and reinforcement learning as key components. For more details, you can refer to their published tech blog, linked here for your reference: https://www.linkedin.com/blog/engineer...