Snacks Weekly on Data Science

In this episode, we explore how Uber tackled one of the most operationally challenging parts of its marketplace: airport pickups. Unlike normal city rides, airport demand is highly bursty, queue-driven, and heavily influenced by flight schedules, delays, and driver positioning decisions. To solve this, Uber built a coordinated forecasting system composed of three specialized models: Estimated Time to Request (ETR) to predict queue wait times, Earnings Per Hour (EPH) to estimate airport profitability versus city driving, and Driver Deficit Forecasting to proactively reposition supply before shortages occur. This allows the platform to reduce uncertainty, improve driver behavior, and stabilize airport marketplace dynamics in real time.

For more details, you can refer to their published tech blog, linked here for your reference: https://www.uber.com/blog/forecasting-models-to-improve-availability-at-airports

In this episode, we discuss Expedia’s need to evaluate many ranking ideas quickly so that A/B testing would not become a bottleneck. We explore their three-stage experimentation funnel: backtesting to remove weak candidates, interleaving to rapidly compare promising ones, and A/B testing to validate final business impact. What made this design approach effective was using the right evaluation tool at the right step instead of forcing A/B testing to do everything.
For more details, you can refer to their published tech blog, linked here for your reference: https://medium.com/expedia-group-tech/interleaving-for-accelerated-testing-75adc644027b

In this episode, we look at how companies deal with large volumes of unstructured text and why traditional clustering methods often fall short at scale. We explore two LLM-powered approaches shared by data scientists from Microsoft: a bottom-up pipeline that builds structure from data using embeddings and clustering, and a top-down pipeline that starts with LLM-generated categories and refines them recursively into a hierarchy.
For more details, you can refer to their published tech blog, linked here for your reference: https://medium.com/data-science-at-microsoft/from-chaos-to-clarity-building-taxonomies-from-unstructured-text-using-large-language-models-c1303db3adb1

In this episode, we discuss a classic scaling problem in fraud and risk operations: too much manual review, inconsistent judgments, and growing complexity. We explore the team’s solution, Bumblebee, a multi-agent AI architecture that separates planning, evidence gathering, and analysis into specialized roles, enabling a robust and scalable system to solve the problem.
For more details, you can refer to their published tech blog, linked here for your reference: https://engineering.razorpay.com/meet-bumblebee-the-multi-agent-ai-architecture-that-changed-fraud-detection-at-razorpay-c2b6d5704f51

In this episode, we explore how OLX improved discovery by combining keyword search and vector search instead of forcing a choice between the two. Keyword systems remain excellent for precision, while vector systems add semantic understanding. Together, they create a smarter and more user-friendly marketplace experience.
For more details, you can refer to their published tech blog, linked here for your reference: https://tech.olx.com/hybrid-search-where-keywords-meet-vectors-enabling-classifieds-discovery-b7c383fe4fc4

In this episode, we explore how Udemy built a multilingual AI platform to bring its generative AI features to learners around the world. The team approached localization across three levels: a translation-first approach for broad and fast coverage, a fully native multilingual system for markets where fluency and cultural precision are essential, and a hybrid solution in between that intelligently routes between the two depending on the situation
For more details, you can refer to their published tech blog, linked here for your reference: https://medium.com/udemy-engineering/from-zero-to-hero-localization-led-generative-ai-at-udemy-a422e4f968d4

In this episode, we explore how Meta uses the “Ladder of Evidence” framework to evaluate the effectiveness of new product features. Instead of relying on a single analytical method, this framework helps teams choose the right type of evidence based on real-world constraints, leading to better and more informed product decisions.

For more details, you can refer to their published tech blog, linked here for your reference: 

https://medium.com/@AnalyticsAtMeta/ladder-of-evidence-in-understanding-effectiveness-of-new-products-part-i-ad8dee70906c

In this episode, we explore how Vimeo built a customized AI system for subtitle translation—one that goes beyond basic text translation to tackle the much more challenging problem of synchronizing language with timing. We discuss how the team designed a split-brain architecture to separate translation quality from timing constraints, and how they implemented fallback mechanisms to ensure the system remains reliable in real-world scenarios.

For more details, you can refer to their published tech blog, linked here for your reference: https://medium.com/vimeo-engineering-blog/how-we-built-ai-powered-subtitles-at-vimeo-ff11f1d64b2a

In this episode, we explore how the New York Times engineering team used AI agents to scale unit test coverage across their News site. They accomplished this by building a custom coverage measurement tool, designing a two-loop human–AI workflow, and investing heavily in prompt engineering, including strict guardrails to prevent the agent from cheating or drifting. The key takeaway is that AI works best when it is tightly constrained, carefully monitored, and used to amplify human judgment. 
For more details, you can refer to their published tech blog, linked here for your reference: https://open.nytimes.com/how-the-new-york-times-is-scaling-unit-test-coverage-using-ai-tools-fa796bf9b8d2

In this episode, we explore how Shopify evolved its product classification system across three major stages: from a traditional logistic regression model with TF-IDF features, to a multi-modal approach combining text and images, and finally to Vision Language Models built on top of a standardized and evolving product taxonomy. We also look at how architectural design and inference optimization are just as important as model accuracy in real-world machine learning systems.
For more details, you can refer to their published tech blog, linked here for your reference: https://shopify.engineering/evolution-product-classification