Episode 28 of The Data Science Podcast with Fexingo explores the growing energy footprint of data centers and AI training. Lucas and Luna break down a real 2025 case: a midwestern utility that had to fast-track two natural gas plants to power a single hyperscale campus. They discuss the gap between model efficiency gains and total compute growth, the role of on-site renewables, and whether the grid can keep up with demand from training runs. Listeners learn why your next model might be constrained not by memory or data, but by the local power substation. Plus, a brief note on how listener support keeps this show ad-free.

#DataCenters #Energy #AITraining #Hyperscale #GridCapacity #PowerDemand #Sustainability #Renewables #NaturalGas #ModelEfficiency #DigitalRealEstate #Technology #DataSciencePodcast #FexingoBusiness #BusinessPodcast #Podcast #DataDriven #Infrastructure

Keep every episode free: buymeacoffee.com/fexingo

Episode 27 of The Data Science Podcast dives into one of the most common yet under-discussed problems in applied machine learning: production data pipeline failures. Lucas and Luna unpack a real-world case from a mid-sized fintech company where a silent schema change broke a credit risk model for three weeks, costing over $2 million in mispriced loans. They walk through the specific failure modes—schema drift, late-arriving data, and upstream dependency breaks—and discuss practical countermeasures like schema enforcement, data quality monitors, and pipeline observability tools. If you've ever had a model silently degrade in production, this episode will give you a framework to diagnose and prevent it.

#DataPipelines #MLOps #ProductionMachineLearning #SchemaDrift #DataQuality #DataObservability #FeatureEngineering #DataEngineering #MachineLearning #Fintech #CreditRiskModeling #PipelineFailures #DataScience #Technology #FexingoBusiness #BusinessPodcast #Podcast #DataSciencePodcast

Keep every episode free: buymeacoffee.com/fexingo

Kaggle competitions have launched data science careers and pushed the field forward, but they also teach habits that don't translate to production. Lucas and Luna break down the gap between leaderboard-chasing and actual data science work—using the iconic 'Porto Seguro's Safe Driver Prediction' competition as a case study. They explore how public leaderboards encourage overfitting, why feature engineering on Kaggle doesn't match enterprise constraints, and what companies like Uber and Airbnb have learned from hiring Kaggle grandmasters. If you've ever wondered whether Kaggle skills prepare you for the real world—or if they can actually hurt your team's models—this episode will change how you think about competition data science.

#Kaggle #DataScience #MachineLearning #Competitions #Overfitting #FeatureEngineering #PortoSeguro #ProductionModels #Leaderboard #Grandmasters #Uber #Airbnb #DataScienceSkills #RealWorldDataScience #Technology #FexingoBusiness #BusinessPodcast #DataDriven

Keep every episode free: buymeacoffee.com/fexingo

Lucas and Luna dive into concept drift—when the statistical properties of a target variable change over time, degrading model performance. Using a concrete case from a credit card fraud detection system, Lucas explains how data scientists monitor for drift using metrics like PSI (Population Stability Index) and KL divergence. Luna challenges whether traditional thresholds are enough, and they explore adaptive strategies like online learning and retraining triggers. Specific examples from e-commerce and finance show how drift detection prevents silent failures. By the end, listeners understand the difference between covariate shift, prior probability shift, and concept shift—and why monitoring is a continuous, not one-time, task.

#DataScience #MachineLearning #ConceptDrift #MLOps #ModelMonitoring #FraudDetection #PopulationStabilityIndex #KullbackLeiblerDivergence #OnlineLearning #CovariateShift #PriorProbabilityShift #ConceptShift #Retraining #FeatureDrift #PredictionDrift #Technology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

Data scientist Lucas and data-savvy Luna dive into data leakage—how a seemingly perfect machine learning model can fail spectacularly because it accidentally learned from information it shouldn't have had. Lucas walks through a concrete example: a hospital's model that predicted sepsis with 99% accuracy, only to crash to 50% in production because it had learned to detect the time-stamp pattern of when lab tests were ordered, not the actual patient condition. Luna pushes on the distinction between target leakage and train-test contamination, and Lucas shares a real case from a credit card fraud detection model that used a 'fraud flag' column that was actually created after the fraud was confirmed. They discuss simple safeguards like time-based splitting and feature provenance tracking. The episode closes with a reflection on why data leakage is the most embarrassing yet instructive failure in data science.

#DataLeakage #MachineLearning #DataScience #ModelFailure #TargetLeakage #TrainTestContamination #FeatureEngineering #FraudDetection #HealthcareAI #SepsisPrediction #ModelValidation #TimeSeriesSplit #FeatureProvenance #DataPipeline #ProductionML #Technology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

Lucas and Luna explore how causal forests — a machine learning method developed from the work of Susan Athey and others — let data scientists estimate ad effectiveness without randomized experiments. Using a real-world case from a mid-market e-commerce company that spent $2 million on YouTube ad campaigns, they show how the company's data team built a causal forest model to isolate the true incremental lift of each campaign, controlling for seasonality, user intent, and demographic confounders. The episode walks through the core idea of heterogeneous treatment effects, the split-criterion differences vs. a standard regression tree, and how the model revealed that one ad creative actually had negative lift for a specific customer segment. No math overload — just the intuition and a concrete result: the company reallocated 40% of its ad budget and saw a 15% increase in return on ad spend the following quarter. The episode also touches on when causal forests beat propensity score matching and A-B tests.

#CausalForests #MachineLearning #DataScience #AdMeasurement #CausalInference #SusanAthey #HeterogeneousTreatmentEffects #MarketingAnalytics #ROI #YouTubeAds #Ecommerce #Technology #BusinessPodcast #FexingoBusiness #DataDriven #AdBudget #IncrementalLift #PropensityScore

Keep every episode free: buymeacoffee.com/fexingo

Episode 22 of The Data Science Podcast dives into a fascinating real-world case: how LinkedIn's data science team used causal inference — specifically a method called double machine learning — to figure out whether tweaking the feed algorithm actually caused more engagement, or if the apparent lift was just correlation. Lucas and Luna walk through the problem of confounding variables, the difference between prediction and causation, and how LinkedIn ran a synthetic control-style analysis on historical data instead of a classic A-B test. Along the way, they discuss the limits of observational causal inference, why LinkedIn chose not to run a live experiment, and what other tech platforms can learn from this approach. If you've ever wondered why 'you might also like' recommendations sometimes feel like noise, this episode gives you the data-science reasoning behind the redesign. Features a brief mention of how listener support via buy me a coffee dot com slash fexingo keeps the show ad-free.

#CausalInference #DoubleMachineLearning #LinkedIn #FeedAlgorithm #EngagementMetrics #DataScience #A-BTesting #ConfoundingVariables #SyntheticControl #TechPlatform #RecommendationSystems #MachineLearning #ObservationalData #Causality #LinkedInLabs #FexingoBusiness #BusinessPodcast #Technology

Keep every episode free: buymeacoffee.com/fexingo

When a data science team grows from two people to twenty, model training turns into chaos. Different engineers pull data from different sources, calculate features differently, and nobody can reproduce a model from six months ago. This episode walks through the real story of a fintech startup that hit exactly this wall: their credit-risk model worked in a notebook but fell apart in production because the feature definitions had silently diverged. Lucas explains how a feature store — a central repository for curated, versioned features — solved the reproducibility crisis and cut model development time by 40 percent. Luna pushes back on whether it's just another infrastructure sink, and they dig into concrete trade-offs: when a feature store helps, when it's overkill, and why the biggest challenge is organizational, not technical. If you've ever wondered why your team's models degrade over time even when the data looks the same, this episode gives you the vocabulary and the architecture to fix it.

#FeatureStore #MachineLearning #DataEngineering #Reproducibility #MLOps #DataScience #FeatureEngineering #ModelGovernance #Fintech #CreditScoring #Tecton #Feast #DataInfrastructure #Tech #Podcast #FexingoBusiness #BusinessPodcast #DataDriven

Keep every episode free: buymeacoffee.com/fexingo

Most data science teams treat documentation as an afterthought—a few comments in a Jupyter notebook or a stale Confluence page. In episode 20 of The Data Science Podcast, Lucas and Luna explore why that approach is dangerous, especially as models move into production. They walk through a real case: a fintech startup whose fraud model silently degraded because feature definitions in the code didn't match the original research notebook. The fix wasn't a better algorithm—it was a living documentation system that versioned assumptions, data schemas, and model cards alongside the code. Lucas explains how tools like DVC and MLflow can enforce this, and why the cost of under-documentation compounds over time. If you've ever inherited a model and wondered 'what was this person thinking?', this episode is for you.

#DataScience #MachineLearning #MLOps #Documentation #ModelGovernance #Reproducibility #MLflow #DVC #FeatureEngineering #DataVersioning #ModelCards #TechPodcast #Technology #DataPipeline #Fintech #FraudDetection #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

Lucas and Luna dive into RLHF — reinforcement learning from human feedback — the technique that made modern chatbots safe and useful. They break down the three-stage pipeline: supervised fine-tuning, reward model training with human comparisons, and proximal policy optimization. Using real examples, they explain why reward hacking is a real danger and how alignment is never truly finished. A must-listen for anyone building or using LLMs.

#RLHF #ReinforcementLearning #HumanFeedback #LLM #ChatGPT #Claude #ProximalPolicyOptimization #RewardHacking #AIAlignment #SupervisedFineTuning #DataScience #Technology #FexingoBusiness #BusinessPodcast #MachineLearning #DeepLearning #AI #GenAI

Keep every episode free: buymeacoffee.com/fexingo