When a model silently fails in production, the root cause often traces back to a forgotten metadata point: which version of the training pipeline was used. In this episode, Lucas and Luna unpack a real case from a fintech lender that lost two million dollars because a model was accidentally retrained on stale features. They walk through the metadata fields every team should track—from git commit hashes to feature store timestamps—and explain how one well-designed training manifest can save weeks of debugging. No abstract theory: this is the nitty-gritty of model governance that separates amateur setups from production-grade data science.

#ModelGovernance #Metadata #MLOps #TrainingPipeline #DataScience #MachineLearning #Fintech #FeatureStore #Reproducibility #ProductionML #DataDrift #VersionControl #Tech #Technology #FexingoBusiness #BusinessPodcast #DataSciencePodcast #TheDataSciencePodcast

Keep every episode free: buymeacoffee.com/fexingo

When fraud detection models flag only 2% of transactions as suspicious, the real criminals often slip through. In this episode, Lucas and Luna unpack how a data scientist at a major payments processor used graph analytics and network features to uncover a hidden fraud ring that had evaded every rule-based and machine learning model for 18 months. The key insight: instead of scoring individual transactions, they built a similarity graph connecting merchants by shared IP addresses, device fingerprints, and refund patterns. A single connected component revealed 47 dummy merchants all feeding into one shell company. The model caught it in week two of deployment, recovering an estimated $340 million in chargebacks over the next quarter. We talk about why classic fraud models miss organised rings, how graph features like PageRank and community detection can surface collusion, and why graph neural networks are becoming the new standard in fintech fraud detection.

#GraphAnalytics #FraudDetection #DataScience #MachineLearning #NetworkAnalysis #Fintech #GraphNeuralNetworks #PageRank #CommunityDetection #AnomalyDetection #PaymentsProcessing #ChargebackFraud #FeatureEngineering #OrganizedFraud #Technology #FexingoBusiness #BusinessPodcast #DataDriven

Keep every episode free: buymeacoffee.com/fexingo

Episode 16 of The Data Science Podcast with Fexingo. Lucas and Luna explore how a major European payments company used synthetic data to fix a fraud detection model that was crippled by privacy regulations and extreme class imbalance. Synthetic data—artificially generated records that preserve statistical patterns without exposing real user information—allowed the team to augment their training set by a factor of 50 and reduce false negatives by 40 percent. The hosts break down the generation technique (a conditional GAN trained on real transaction metadata), discuss the evaluation challenge (how do you know the synthetic data is safe?), and consider the broader trend of 'data sovereignty' in machine learning. A practical episode for data scientists wrestling with thin or sensitive datasets.

#SyntheticData #FraudDetection #GAN #DataAugmentation #Privacy #MachineLearning #ClassImbalance #ModelValidation #Fintech #PaymentsIndustry #ConditionalGAN #DataSovereignty #EnterpriseML #Technology #DataSciencePodcast #FexingoBusiness #BusinessPodcast #ProductionML

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna dive into the collaborative filtering algorithm behind Spotify's Discover Weekly. They break down how matrix factorization learns user preferences from sparse listening data, using the example of a user who jumps from classical to indie rock. They discuss cold-start problems, the trade-off between exploration and exploitation, and why your weekly playlist feels eerily accurate. Along the way, they touch on implicit feedback signals like skip rate and repeat listens, and compare Spotify's approach to those of Pandora and Apple Music. If you've ever wondered how a machine can predict your next favorite song, this episode explains the math and the engineering choices that make it work.

#Spotify #DiscoverWeekly #CollaborativeFiltering #MatrixFactorization #RecommendationSystems #MachineLearning #Personalization #ColdStartProblem #ExplorationVsExploitation #ImplicitFeedback #DataScience #Technology #FexingoBusiness #BusinessPodcast #Podcast #MusicRecommendation #UserBehavior #Algorithms

Keep every episode free: buymeacoffee.com/fexingo

Ever wonder how Spotify knows what song you want to hear next? This episode unpacks the collaborative filtering algorithm that powers Discover Weekly and Release Radar. Lucas walks through the math behind user-item matrices, matrix factorization, and how Spotify tackles the cold-start problem for new users and new artists. Luna probes the trade-off between personalization and the filter bubble, and the hosts discuss why Spotify's A-B testing showed that a 70/30 blend of familiar and novel tracks maximizes user satisfaction. Specific data: how the 'Taste Profile' matrix handles 100 million active users and 70 million tracks, and why the algorithm still recommends a song you already skipped.

#Spotify #CollaborativeFiltering #RecommendationEngine #MatrixFactorization #DiscoverWeekly #ColdStartProblem #FilterBubble #ABTesting #UserItemMatrix #Personalization #DataScience #MachineLearning #Technology #FexingoBusiness #BusinessPodcast #DataDriven #LucasAndLuna #TheDataSciencePodcast

Keep every episode free: buymeacoffee.com/fexingo

Episode 13 of The Data Science Podcast with Fexingo dives into causal inference—specifically, how data scientists can estimate cause-and-effect relationships from observational data when A-B testing isn't possible. Lucas and Luna walk through a real-world case: how a health-tech startup used double machine learning (DML) to determine whether its app's push notifications actually reduced hospital readmissions, without running a randomized trial. They break down the core challenge—confounding variables—and explain how DML uses machine learning to model both the treatment and the outcome, then isolates the causal effect. The conversation covers the 'honest' approach of sample splitting to avoid overfitting bias, and why this method is gaining traction in fields like economics, epidemiology, and marketing. By the end, listeners will understand the difference between correlation and causation in a practical, code-adjacent way, and know one concrete technique to try when an A-B test is off the table.

#CausalInference #DoubleMachineLearning #DataScience #ObservationalData #ConfoundingVariables #A-BTesting #CausalEffect #MachineLearning #HealthcareAnalytics #ReadmissionRates #PushNotifications #TreatmentEffect #SampleSplitting #Econometrics #Technology #FexingoBusiness #BusinessPodcast #DataDriven

Keep every episode free: buymeacoffee.com/fexingo

Episode 12 of The Data Science Podcast dives into why traditional frequentist A/B testing can lead to false positives and how a Bayesian approach fixes it. Lucas and Luna walk through a concrete example: an e-commerce team testing a new checkout flow that looked like a winner at 5,000 visitors but collapsed at 10,000. They explain p-hacking, the peek problem, and how Bayesian methods with prior distributions, posterior probabilities, and expected loss give you more reliable decisions with smaller sample sizes. No math PhD required — just practical intuition for data scientists and product managers who run tests every week. This episode also covers the free, ad-free mission of the podcast and how listeners can support it.

#BayesianA-BTesting #FrequentistStatistics #p-value #p-hacking #DataScience #A-BTesting #BayesianInference #PriorDistribution #PosteriorProbability #ExpectedLoss #FalsePositive #EcommerceTesting #ProductManagement #Experimentation #MachineLearning #StatisticalMethods #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

Episode 11 of The Data Science Podcast tackles the hidden danger of imbalanced datasets. Lucas and Luna walk through a real-world example: a fraud detection model trained on 99.9 percent legitimate transactions and 0.1 percent frauds. The model achieved 99.9 percent accuracy yet caught zero frauds. They explain why accuracy is a terrible metric on imbalanced data, introduce precision-recall curves and F1-score as better alternatives, and discuss resampling techniques like SMOTE and cost-sensitive learning. Listeners will learn how to spot imbalance traps in their own projects and why some problems require rethinking the loss function entirely. The conversation stays practical and code-adjacent without getting lost in syntax. If you have ever trained a classifier on skewed data and felt something was off, this episode will give you the diagnostic tools to fix it.

#ImbalancedData #Classification #FraudDetection #PrecisionRecall #F1Score #SMOTE #CostSensitiveLearning #DataScience #MachineLearning #ModelEvaluation #AccuracyTrap #Resampling #ClassImbalance #Technology #BusinessPodcast #FexingoBusiness #TheDataSciencePodcast #Fexingo

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna tackle one of the most frustrating problems in modern AI: hallucination in large language models. They break down the specific mechanisms that cause models to confidently generate false information, using the example of a customer support chatbot that invented a refund policy. Lucas explains how retrieval-augmented generation (RAG) and grounding techniques can reduce hallucination rates from over 20 percent to under 5 percent, citing a 2025 paper from Google DeepMind. They also discuss trade-offs with latency and cost, and why no approach is perfect yet. The conversation stays grounded in real numbers and concrete engineering decisions, giving listeners a clear framework for diagnosing and mitigating hallucinations in their own applications.

#Hallucination #LargeLanguageModels #RAG #RetrievalAugmentedGeneration #AIModels #MachineLearning #NLP #Chatbot #PromptEngineering #Grounding #GoogleDeepMind #AIAccuracy #DataScience #Technology #FexingoBusiness #BusinessPodcast #ModelReliability #AISafety

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna explore how interpretability tools like SHAP and LIME uncovered a hidden signal in a hospital's cancer diagnosis model. They walk through a real case where the model was accurate but biased by a spurious correlation with patient age, and how a data scientist used local explanations to catch it before deployment. The conversation covers the difference between global and local interpretability, why accuracy metrics can mask dangerous blind spots, and how one hospital saved lives by asking 'why' instead of just 'how good'. Lucas and Luna also touch on the trade-off between model complexity and explainability, and why regulators are starting to demand interpretable models in healthcare. Listeners will come away with a concrete example of why interpretability isn't just a nice-to-have but a critical safety check in high-stakes machine learning.

#InterpretableML #SHAP #LIME #HealthcareAI #CancerDiagnosis #ModelBias #DataScience #MachineLearning #Explainability #FeatureImportance #ModelValidation #AIEthics #ClinicalAI #DataDriven #Podcast #Technology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo