エピソード

  • How MLOps Teams Are Using Model Monitoring to Prevent Silent Failures

    How MLOps Teams Are Using Model Monitoring to Prevent Silent Failures
    2026/06/08
    Episode 39 of The Data Science Podcast explores the growing discipline of model monitoring in production. Lucas and Luna discuss why many data science teams still treat monitoring as an afterthought, how silent failures erode business trust, and what tools like Evidently AI, WhyLabs, and custom dashboards are doing about it. They walk through a real example from a fintech lending platform where a model drifted undetected for weeks, costing the company over $2 million in bad loans. The conversation also covers the three key pillars of monitoring: data quality, model performance, and operational health. Lucas shares a practical checklist for teams getting started with monitoring today. If you are deploying models to production, this episode will save you from waking up to a 3 AM pager alert. #ModelMonitoring #MLOps #DataScience #MachineLearning #ProductionML #SilentFailures #ConceptDrift #DataQuality #MLPipeline #Fintech #EvidentlyAI #WhyLabs #MLObservability #DataDrift #ModelGovernance #FexingoBusiness #BusinessPodcast #Technology Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    9 分
  • How Data Scientists Use Bayesian A-B Testing

    How Data Scientists Use Bayesian A-B Testing
    2026/06/08
    Lucas and Luna dive into Bayesian A/B testing, a method that's quietly replacing traditional frequentist approaches in data science. They break down how it works, why it's more intuitive, and where it falls short. The episode centers on a real case: how a major retailer used Bayesian testing to optimize their checkout flow, cutting decision time from weeks to days. Lucas explains the math behind prior probabilities and posterior distributions without the jargon, while Luna questions whether Bayesian methods can really scale in big-tech environments. They also touch on the common pitfalls, like choosing a bad prior or misinterpreting results. By the end, listeners will understand the key difference between 'is this statistically significant?' and 'what's the probability this variant is better?'—and why the latter question is often more useful in practice. #DataScience #Technology #BayesianStatistics #ABTesting #MachineLearning #StatisticalModeling #DataDrivenDecisionMaking #PriorProbability #PosteriorDistribution #ConversionRateOptimization #FrequentistVsBayesian #DataSciencePodcast #FexingoBusiness #BusinessPodcast #Experimentation #DecisionScience #EcommerceAnalytics #DataCulture Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    8 分
  • How Spotify Uses Data to Predict Your Next Favorite Song

    How Spotify Uses Data to Predict Your Next Favorite Song
    2026/06/07
    In this episode of The Data Science Podcast, Lucas and Luna dive into Spotify's recommendation engine — not the playlist curation you already know, but the predictive models that identify tracks you haven't heard yet. They break down the specific machine learning techniques behind Spotify's 'Discover Weekly' and 'Release Radar,' focusing on collaborative filtering, natural language processing of audio features, and the bandit algorithms that balance exploration and exploitation. Lucas explains how Spotify processes over 30 billion listening events per day to train its models, and why the company uses a two-tower neural network architecture for candidate generation and ranking. Luna asks the tough questions: how does Spotify avoid filter bubbles, and what happens when the model recommends a song you hate? They also touch on the ethical considerations of hyper-personalization and the trade-offs between user satisfaction and data collection. If you've ever wondered how an algorithm knows your musical taste better than you do, this episode delivers a concrete, behind-the-scenes look at one of the most sophisticated recommender systems in production today. #Spotify #RecommendationSystem #CollaborativeFiltering #MachineLearning #DataScience #NeuralNetworks #BanditAlgorithms #Personalization #MusicDiscovery #AudioFeatures #TwoTowerModel #FilterBubble #Tech #DataDriven #FexingoBusiness #BusinessPodcast #DataSciencePodcast #Fexingo Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    10 分
  • How Netflix Uses Bandit Algorithms for Thumbnail Selection

    How Netflix Uses Bandit Algorithms for Thumbnail Selection
    2026/06/07
    Netflix tests millions of thumbnail combinations daily using multi-armed bandit algorithms — a technique smarter than traditional A/B testing. In this episode, Lucas and Luna break down how contextual bandits balance exploration and exploitation to serve the most clickable image for each user, without wasting traffic on losing variants. They walk through the epsilon-greedy approach, why Netflix abandoned pure A/B splits, and how this same logic applies to ad creatives, recommendation carousels, and any high-stakes personalisation problem. Real example: how a single show's artwork variant drove a 20% lift in plays. If you work with user-facing models or optimise for engagement metrics, this episode gives you a practical framework you can borrow tomorrow. #Netflix #BanditAlgorithms #MultiArmedBandit #ContextualBandit #ThumbnailPersonalization #A/BTesting #ExplorationVsExploitation #EpsilonGreedy #RecommendationSystems #ConversionOptimization #DataScience #MachineLearning #Personalization #Technology #FexingoBusiness #BusinessPodcast #DataDriven #UserEngagement Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    12 分
  • How Data Scientists Measure Model Fairness in Practice

    How Data Scientists Measure Model Fairness in Practice
    2026/06/06
    Episode 35 of The Data Science Podcast tackles the messy reality of algorithmic fairness. Lucas and Luna explore why standard fairness metrics can conflict, using the real-world example of a lending model that passed statistical parity tests but failed an individual fairness audit. They discuss the Impossibility Theorem of Fairness, how researchers at MIT measured bias in commercial facial recognition, and why fairness is a product decision, not a mathematical one. The hosts also share practical steps data scientists can take today: documenting design choices, running disaggregated evaluations, and building simple guardrail dashboards. No platitudes, just the trade-offs and tools that practitioners actually use. #AlgorithmicFairness #ModelBias #FairnessMetrics #DataScience #MachineLearning #LendingModel #FacialRecognition #ImpossibilityTheorem #DisaggregatedEvaluation #GuardrailDashboard #EthicsInAI #TechEthics #DataSciencePodcast #FexingoBusiness #BusinessPodcast #TechnologyPodcast #Podcast #Fexingo Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    6 分
  • How Data Scientists Use Synthetic Data to Beat Data Scarcity

    How Data Scientists Use Synthetic Data to Beat Data Scarcity
    2026/06/06
    When there's not enough real data to train a model, data scientists are turning to synthetic data — artificial datasets generated from a small sample of real observations. In this episode, Lucas and Luna unpack how a healthcare startup used synthetic data to train a rare-disease diagnostic model when only 200 real patient records existed. They walk through the generation techniques — from simple bootstrapping to GANs and diffusion models — and the hidden risk of 'synthetic bias' where artifacts in generated data fool the model. The episode also covers the open-source libraries turning synthetic data from a research trick into a production tool, and why regulators are starting to pay attention. A concrete look at the practice that lets data scientists do more with less. #SyntheticData #DataScarcity #GenerativeAI #GANs #DiffusionModels #HealthcareAI #RareDisease #MachineLearning #DataScience #MLOps #BiasInAI #DataAugmentation #OpenSource #SDV #Technology #BusinessPodcast #FexingoBusiness #TheDataSciencePodcast Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    10 分
  • How Data Scientists Use Causal Inference to Measure Marketing ROI

    How Data Scientists Use Causal Inference to Measure Marketing ROI
    2026/06/05
    In Episode 33 of The Data Science Podcast, Lucas and Luna drill into a single question that has vexed marketers and analysts alike for decades: how do you know if an ad really caused a sale, or if the person would have bought anyway? Rather than a theoretical overview, the episode centers on a concrete case from a mid-size e-commerce company that ran a geo-level experiment poking holes in its own attribution model. The hosts walk through the difference between correlation and causation in a marketing context, explain why last-click attribution is dangerously misleading, and show how a simple difference-in-differences design revealed that the company's most expensive ad channel was actually cannibalizing organic demand. Along the way, Lucas and Luna debate the practical trade-offs of randomized experiments vs. quasi-experimental methods, and share specific metrics data scientists should track before signing off on a marketing-mix model. The episode closes with a forward-looking question about whether AI-generated content will force a new generation of causal tools. #CausalInference #MarketingAnalytics #DataScience #AttributionModeling #DifferenceInDifferences #ADPilot #MarketingMixModeling #EcommerceAnalytics #DataDrivenMarketing #LastClickAttribution #AITools #ExperimentDesign #GeoExperiment #ConversionRate #ROI #BusinessPodcast #FexingoBusiness #Technology Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    10 分
  • How Data Scientists Automate Model Retraining

    How Data Scientists Automate Model Retraining
    2026/06/05
    Episode 32 of The Data Science Podcast explores the critical difference between deploying a model and keeping it relevant. Lucas and Luna break down why most ML teams treat model retraining as a fire drill instead of a scheduled process. They walk through the real-world case of an e-commerce company whose recommendation engine silently deteriorated over six months because no one monitored feature drift. The hosts explain practical patterns: scheduled retraining vs. trigger-based retraining, why locking training data snapshots prevents reproducibility disasters, and how one team used a simple dashboard to catch a 12 percent accuracy drop before it hit revenue. They also discuss the human side: why data scientists resist automation, and how the industry is slowly moving toward MLOps standards that treat model maintenance like software maintenance. A concrete, actionable episode for anyone who has ever wondered what happens after the model goes into production. #ModelRetraining #MLOps #FeatureDrift #DataScience #MachineLearning #ModelMaintenance #ProductionML #Reproducibility #Ecommerce #RecommendationEngine #DataPipeline #Automation #Technology #FexingoBusiness #BusinessPodcast #TheDataSciencePodcast #LucasAndLuna #DataTeam Keep every episode free: buymeacoffee.com/fexingo
    続きを読む 一部表示
    13 分