Episodes
Why Your Data Models Fail at Scale
Most data teams build models that work perfectly in the lab but collapse under real-world traffic. In this episode, we explore how Microsoft tackled massive latency issues by moving from batch processing to online learning systems. We break down the specific trade-offs between model freshness and computational cost, and why the ten percent improvement in response time mattered more than any accuracy metric. Lucas and Luna dissect the engineering decisions behind scalable AI infrastructure.…
How Data Teams Measure AI Impact Beyond Accuracy
Most data teams celebrate model accuracy while ignoring whether the model actually moves business needles. Lucas and Luna drill into why optimizing for F-score or RMSE often leads to production paralysis. They examine a specific case where a major retail chain shifted from predictive precision to causal uplift modeling, revealing how true ROI emerges only when we measure decision quality rather than prediction quality. This episode offers a concrete framework for aligning machine learning…
How Data Teams Build Causal AI Models
Most data teams are stuck in correlation. They build models that predict what happens but can’t explain why, leading to costly mistakes when interventions change the underlying environment. In this episode, Lucas and Luna explore how forward-thinking organizations are shifting from predictive machine learning to causal inference. We look at a specific case where a major logistics provider used do-calculus and structural equation modeling to distinguish between weather-driven demand spikes and…
How Data Teams Build Model Cards for Transparency
In this episode of The Data Science Podcast with Fexingo, Lucas and Luna explore the emerging practice of model cards. They examine how leading technology teams are using standardized documentation to disclose a machine learning system’s intended use, performance metrics across demographic groups, and known limitations. Rather than treating models as black boxes, data science teams are adopting transparency frameworks similar to nutrition labels to build trust with regulators and end users. The…
How Data Teams Build Explainable AI Systems
We explore why black box models are failing enterprise trust and how data teams are shifting toward inherently interpretable architectures. Using a specific case from a major fintech lender, we look at the trade-off between raw predictive power and regulatory compliance. The episode breaks down three practical techniques for building models that explain themselves without sacrificing accuracy. #ExplainableAI #XAI #ModelInterpretability #DataScience #MachineLearning #FexingoBusiness…
How Data Teams Measure ROI Beyond Accuracy
We explore the hidden gap between model accuracy and actual business value, using a specific case where a retail giant’s high-precision inventory model failed to move product. Lucas and Luna break down why optimizing for F1 scores can lead to zero return on investment, and how leading data teams are shifting toward causal inference and marginal lift modeling to prove true economic impact. This is not about better algorithms, it is about better accounting. #DataScience #MachineLearning…
How Data Teams Handle Concept Drift
In this episode of The Data Science Podcast with Fexingo, Lucas and Luna dive into the subtle but critical issue of concept drift. While feature drift is well understood, concept drift represents a fundamental shift in the relationship between input data and target variables over time. Using examples from e-commerce recommendation engines and credit scoring models, we explore why model accuracy can silently degrade even when input distributions remain stable. The hosts discuss detection…
How Data Teams Use Synthetic Data to Train AI
Real data is messy, biased, and expensive. Synthetic data offers a way to generate realistic training sets without touching private information. We look at how major firms are using generative models to create artificial datasets for healthcare and finance, and whether this shortcut actually works or just moves the bias elsewhere. This episode explores the mechanics of synthetic data generation, its applications in privacy-preserving analytics, and the risks of training on fabricated reality.…
Why Your AI Models Are Failing in Production
We explore the hidden gap between model accuracy and business value using a specific case study from a major fintech lender. Discover why optimizing for precision creates silent losses, how to measure true economic impact, and the practical framework data teams use to align algorithmic performance with real-world margins. This episode breaks down the difference between statistical fidelity and financial utility. #FexingoBusiness #BusinessPodcast #DataScience #MachineLearning #ModelMonitoring…
How Data Teams Measure Model Impact Beyond Accuracy
Most data teams celebrate high accuracy scores but fail to track whether those models actually move the business needle. In this episode, Lucas and Luna dissect the gap between technical performance and commercial value using a specific retail inventory case where a model improved precision by two percent yet reduced overall profitability due to ignored opportunity costs. They introduce the concept of decision-centric evaluation, showing how to map prediction errors directly to P&L impact…
How Data Teams Use Data Lineage to Trace Model Decisions
We talk about the hidden cost of black-box models and why tracing data lineage is no longer optional for regulated industries. Using a recent compliance failure at a major fintech firm as our anchor, we break down how lineage graphs connect raw inputs to final predictions. We discuss the practical tools like OpenLineage and Marquez, and why tracking metadata matters more than just tracking model accuracy. If you are building AI systems that affect people's lives, this episode explains how to…
How Data Teams Use Feature Engineering to Build Better Models
We talk about the often overlooked art of feature engineering and why raw data rarely wins on its own. Lucas and Luna break down how transforming variables, like creating interaction terms or encoding categorical noise, can double model performance without touching a single line of deep learning code. We look at a specific case from a mid-sized retail analytics team that improved their churn prediction by thirty percent simply by restructuring their time-series features. It is a reminder that…
How Data Teams Use Survival Analysis to Predict Churn
In this episode of The Data Science Podcast, Lucas and Luna explore how data teams apply survival analysis to predict customer churn. They unpack the core concepts—hazard functions, Kaplan-Meier curves, and Cox proportional hazards—using a relatable example: a subscription streaming service losing customers at different rates. They discuss why traditional classification models often miss the timing of churn, how censoring works, and how survival models can guide retention strategies like…
How Data Scientists Use Knowledge Distillation to Compress Models
In episode 170 of The Data Science Podcast, Lucas and Luna explore knowledge distillation—a technique for compressing large, powerful machine learning models into smaller, faster ones. They use the concrete example of a large language model distilled into a compact version that runs on a smartphone, cutting latency and energy use while preserving most of the accuracy. The hosts explain the core idea: a 'teacher' model trains a 'student' model to mimic its predictions, transferring knowledge…
How Data Teams Use Shapley Values to Explain Model Predictions
In episode 169 of The Data Science Podcast, Lucas and Luna explore how data scientists use Shapley values to explain individual model predictions. They break down the cooperative game theory behind the method, walk through a real-world example from credit lending, and discuss the practical trade-offs between accuracy and interpretability. The hosts explain why Shapley values are the gold standard for feature attribution, how they differ from simpler methods like LIME, and how modern tools like…
How Data Teams Use Data Contracts to Prevent Pipeline Chaos
Episode 168 of The Data Science Podcast with Fexingo dives into data contracts — the agreements that keep data pipelines from turning into chaos. Lucas and Luna explore how teams at companies like Uber and Airbnb use schema agreements to catch breaking changes before they hit production, referencing the 2023 incident at a major bank where a missing contract caused a costly outage. They explain the core components of a data contract, the rollout challenges, and the cultural shift required to…
How Data Teams Use Embedding Drift to Catch Model Decay
In this episode of The Data Science Podcast, Lucas and Luna explore how data teams use embedding drift to detect when machine learning models silently decay in production. They start with a concrete example: a fraud model at a large bank that started missing new scam patterns after a year in production. The conversation covers what embeddings are, why monitoring drift in embedding space is more sensitive than tracking raw feature distributions, and practical methods like cosine similarity…
How Data Teams Use Embeddings to Capture Meaning
In this episode, Lucas and Luna explore how data teams use embeddings to turn text, images, and users into vectors that machines can reason over. They start with the concrete example of Airbnb's 2018 shift to embedding-based search, which boosted bookings by nearly five percent, and walk through why representing meaning as coordinates in high-dimensional space is so powerful. They discuss the trade-offs between static embeddings like Word2Vec and contextual models like BERT, how to evaluate…
How Data Scientists Use Probabilistic Programming for Robust Forecasting
In this episode of The Data Science Podcast, Lucas and Luna explore probabilistic programming as a practical tool for modern data teams. Using the example of a fictional e-commerce company forecasting holiday sales, they show how probabilistic models like PyMC and Stan produce full distributions rather than point estimates, giving businesses a clearer view of uncertainty. They discuss concepts like prior distributions, Bayesian inference, and Markov Chain Monte Carlo methods, and how these…
How Data Teams Use Incremental Learning for Streaming Data
Data moves fast. In this episode of The Data Science Podcast, Lucas and Luna explore how data teams use incremental learning to keep models fresh without retraining from scratch every day. They walk through the concept of online learning, contrast it with batch retraining, and dig into a concrete example: a fraud-detection system that updates in near real time to catch new scam patterns as they emerge. They discuss practical techniques like stochastic gradient descent updates, the trade-offs…
How Data Scientists Use Feature Selection to Cut Noise
In this episode of The Data Science Podcast, Lucas and Luna dive into feature selection—the art of choosing which variables actually matter for your model. They explore why more data isn't always better, how redundant features can quietly degrade performance, and the practical techniques data scientists use to trim the fat: from filter methods like correlation and mutual information, to wrapper methods like recursive feature elimination, and embedded approaches like Lasso and feature…
How Data Scientists Use Counterfactual Explanations
In Episode 162 of The Data Science Podcast, Lucas and Luna explore counterfactual explanations—a technique that answers 'what would need to change for this model's decision to flip?' They anchor the discussion in a real-world example: a loan denial from a fintech app. The episode breaks down how counterfactuals differ from feature importance, why they matter for building trust and debugging models, and how they're generated using nearest-neighbor search and optimization. Lucas and Luna also…
How Data Teams Use Model Monitoring to Catch Silent Failures
In this episode of The Data Science Podcast, Lucas and Luna dive into the world of model monitoring — the practice of keeping machine learning models honest after they're deployed. Using the concrete example of a credit-card fraud detection system, they explain why accuracy can stay high while the model quietly goes stale, and how metrics like drift detection, data quality checks, and prediction distribution shifts help catch problems early. They also discuss the difference between monitoring…
How Data Scientists Use Transfer Learning to Save Time
In this episode of The Data Science Podcast, Lucas and Luna explore transfer learning—a technique that lets data scientists reuse pre-trained models instead of starting from scratch. They anchor the discussion with a concrete example: how a small e-commerce team adapted a pre-trained language model to their niche product reviews, cutting training time from weeks to days and data requirements by over 80 percent. They break down the difference between feature extraction and fine-tuning, when…
How Data Teams Use Feature Stores to Keep Models Honest
Feature stores have quietly become the backbone of modern machine learning, but they're not just a place to stash tables. In this episode, Lucas and Luna dig into why feature stores are the difference between a model that works in the lab and one that works in production. They break down the concrete mechanics—point-in-time correctness, training-serving skew, and feature reuse across teams—and walk through a real example from a fintech that cut its model retraining time from weeks to hours.…
How Data Scientists Use Active Learning to Cut Labeling Costs
Labeling data is one of the most expensive bottlenecks in machine learning, but active learning offers a smarter path. In this episode, Lucas and Luna break down how data scientists use active learning to train high-performing models with a fraction of the labeled data. They walk through the key strategies—uncertainty sampling, query-by-committee, and expected model change—and explain why the approach is especially powerful for niche domains like medical imaging and rare-event detection.…
How Data Scientists Use Conformal Prediction for Reliable AI
In this episode, Lucas and Luna unpack conformal prediction, a method that gives machine learning models a rigorous way to state their own uncertainty. They walk through a real-world example—a hospital risk score that must decide when to say 'I don't know.' Along the way, they compare conformal prediction to Bayesian methods, discuss why calibration alone isn't enough, and look at how it's showing up in production systems from self-driving cars to fraud detection. If you've ever wondered how to…
How Data Scientists Use Multi-Armed Bandits to Balance Exploration and Exploitation
In this episode, Lucas and Luna explore the multi-armed bandit problem, a classic dilemma in decision-making under uncertainty that has found new life in data science. They break down how companies like Netflix and Amazon use bandit algorithms to dynamically balance exploration (trying new options) and exploitation (pushing known winners), all without the heavy cost of traditional A/B testing. Using the fictional example of a music streaming service, they walk through the mechanics of…
How Data Scientists Use Manifold Learning to Untangle High-Dimensional Data
In this episode, Lucas and Luna explore manifold learning, a family of techniques that lets data scientists see the true shape of high-dimensional data by projecting it into lower dimensions. Using the classic example of the Swiss roll dataset and real-world applications like facial recognition and sensor diagnostics, they explain how methods like t-SNE and UMAP work, what they get right, and where they can mislead. Along the way, they discuss the curse of dimensionality, how to choose between…
How Data Scientists Use Synthetic Data to Bridge Privacy and Utility
Synthetic data is quietly becoming one of the most practical tools in a data scientist's kit — letting teams train models on realistic datasets without exposing sensitive records. In this episode, Lucas and Luna dig into how synthetic data works, where it shines, and where it still stumbles. They walk through a concrete example: a healthcare startup using synthetic patient records to build a predictive model for hospital readmissions, cutting data-sharing friction with partner institutions…
How Data Scientists Use Causal Forests for Personalized Treatment Effects
In this episode, Lucas and Luna explore how data scientists are using causal forests to estimate personalized treatment effects—moving beyond average outcomes to understand which customers, patients, or users respond best to a specific intervention. They dive into a real-world example: a financial services firm testing a new engagement email, where the causal forest reveals that the average uplift hides a wide variance—some segments respond strongly, others see zero effect. The conversation…
How Data Scientists Use Quantile Regression for Uncertainty
Episode 152 of The Data Science Podcast dives into quantile regression, a technique that gives data scientists a fuller picture of uncertainty than standard mean-focused models. Lucas and Luna explore how predicting the 90th percentile instead of the average can transform decision-making in finance, healthcare, and logistics. They walk through a concrete example from e-commerce, comparing quantile regression to ordinary least squares, and discuss practical implementation using libraries like…
How Data Scientists Use Gradient Boosting for Smarter Predictions
In this episode, Lucas and Luna dive into gradient boosting, the machine learning technique behind countless winning models on Kaggle and in production systems. They explore how it works, why it's so effective, and the practical trade-offs teams face when using it. Using a concrete example of a churn prediction model at a telecom company, they break down the key hyperparameters like learning rate and tree depth, and explain how to avoid the classic pitfall of overfitting. They also discuss the…
How Data Scientists Use Calibration to Fix Model Probabilities
In this milestone 150th episode, Lucas and Luna dive into probability calibration — the practice that makes a model's predicted confidence match real-world outcomes. They anchor the discussion with a concrete example: a fraud model that says '88 percent suspicious' but is right only 65 percent of the time. They explain why raw model outputs can be overconfident, especially in deep learning, and walk through practical fixes like Platt scaling and isotonic regression. Along the way, they touch on…
How Data Scientists Use Data Lineage to Debug Pipelines
When a model's predictions start drifting from reality, the first question every data scientist asks is: where did the data go wrong? In this episode, Lucas and Luna explore a subtle but powerful debugging technique: data lineage. They walk through a real example at a mid-sized e-commerce company where a sudden drop in recommendation click-through rates turned out to trace back to a silent upstream schema change in a product catalog feed. Lucas explains how lineage graphs — the network of data…
How Data Scientists Use Knowledge Graphs to Power Recommendation Engines
In this episode of The Data Science Podcast, Lucas and Luna explore how knowledge graphs are transforming recommendation engines beyond collaborative filtering. They dive into a real example: how a music streaming service uses knowledge graphs to recommend obscure tracks that collaborative filtering would miss. The hosts break down the graph structure, the role of graph embeddings, and how hybrid approaches combine the best of both worlds. They also discuss the practical challenges of building…
How Data Scientists Use Shapley Values for Fairer Models
Lucas and Luna explore Shapley values, the game-theory concept behind SHAP, which data scientists use to fairly distribute model predictions among features. They break down the math in plain English, show why it's more principled than simple feature importance, and discuss its real-world impact on model fairness and interpretability. With examples from credit scoring and fraud detection, they explain how Shapley values help data teams answer 'why did the model say no?' and why that matters for…
How Data Scientists Use Anomaly Detection in Manufacturing
On this episode, Lucas and Luna drill into anomaly detection in manufacturing, using the semiconductor industry as a case study. They explore how a fab uses sensor data from etching tools to spot subtle process drifts before they ruin a batch of wafers — saving millions in scrap and downtime. The conversation covers the shift from simple threshold alerts to unsupervised and semi-supervised models, the challenge of labeling rare defects, and how domain engineers and data scientists collaborate…
How Data Scientists Use Embeddings for Search
Lucas and Luna explore how data scientists are using embeddings to transform search. From semantic understanding to handling typos and synonyms, they unpack the shift from keyword matching to meaning-based retrieval. With a practical case study on Wikipedia's search overhaul, they discuss the challenges of scaling embeddings, the importance of evaluation, and how generative AI is changing the game. If you've ever wondered what powers the search bar in your favorite app, this episode breaks it…
How Data Scientists Reduce Model Bias with Fairness Metrics
Bias in machine learning isn't just a fairness problem — it's a business risk. In this episode, Lucas and Luna unpack how data scientists are using fairness metrics like demographic parity and equalized odds to audit models before they ship. They walk through a real case: a credit approval model that quietly denied loans to applicants from certain postal codes, and how the team caught it with a bias audit. Along the way, they discuss the limits of fairness metrics, the trade-off between…
Why Your Recommendation Model Needs Online Learning
Recommender systems quietly power everything from your streaming queue to your shopping feed, but most teams still retrain them on a fixed schedule. In this episode, Lucas and Luna dig into online learning — the technique that lets models update continuously as new user behavior streams in. They break down the difference between batch retraining and incremental updates, and walk through a real-world case: a mid-size e-commerce company that cut its model update lag from once a night to every few…
Data Drift: How Models Silently Decay and What Teams Do About It
In this episode, Lucas and Luna unpack the quiet killer of machine learning systems: data drift. They trace how models trained on 2024 data start stumbling by 2026, using a concrete example from retail demand forecasting where a model quietly degraded over eighteen months. The conversation walks through the difference between covariate shift and concept drift, why traditional retraining schedules often miss the real problem, and how modern ML teams use statistical tests like the…
How Data Scientists Use Counterfactual Reasoning on Campaigns
In episode 141, Lucas and Luna dig into counterfactual reasoning in data science — the art of asking what would have happened if a marketing campaign never ran. They open with a concrete example: a national grocery chain that used a synthetic control to measure a loyalty program's true lift. The episode walks through why simple before-and-after comparisons overstate results, how to build a credible counterfactual with donor pools and matching, and where this reasoning breaks down in real-world…
How Data Scientists Use Survival Analysis for Customer Lifetime Value
Survival analysis, a statistical method from epidemiology, is becoming a powerful tool for predicting customer churn and lifetime value. In this episode, Lucas and Luna explain how hazard functions and Kaplan-Meier curves give businesses a richer view of retention than traditional classification models. They walk through a real example of a music streaming service using Cox proportional hazards to identify key churn drivers—like playlist frequency and subscription length—months before…
How Data Contracts Are Fixing Broken Data Pipelines
Data contracts are emerging as a powerful pattern to ensure data reliability and quality across modern data stacks. In this episode, we explore how a large European e-commerce platform reduced pipeline failures by 42 percent by implementing data contracts using Great Expectations and CI/CD integration. We discuss what data contracts are, how they differ from simple documentation, and why both small teams and enterprise data organizations are adopting them. Learn how ML teams are using contracts…
How Data Scientists Use Causal Inference for Marketing Attribution
Most marketing attribution models are fundamentally broken — they credit the last click or spread credit evenly, but they don't measure true incrementality. In this episode, Lucas and Luna look at how a major online retailer used causal inference — specifically, a difference-in-differences approach paired with a synthetic control — to measure the actual lift of their email campaigns. They walk through the data requirements, the model architecture (a lightweight gradient boosting model with a…
How Data Scientists Use Feature Stores for Consistent ML Pipelines
In episode 137 of The Data Science Podcast, Lucas and Luna dive into feature stores—a critical piece of MLOps infrastructure that saves data scientists from reinventing the wheel every time they build a model. Lucas cites a 2023 survey showing that data scientists spend 72% of their time on data preparation and feature engineering, not modeling. He explains how feature stores serve as a central repository for versioned, reusable features that work for both training and real-time inference.…
How Bayesian A/B Testing Speeds Up Experimentation
In this episode, Lucas and Luna explore how leading tech companies are shifting from frequentist to Bayesian A/B testing to reduce experiment duration and make better decisions with smaller sample sizes. They walk through the key concepts — prior and posterior distributions, the probability of being best, and practical choices like using a Beta-Binomial model. Lucas draws on a real-world example from a major music streaming platform to show how Bayesian methods can cut experiment time by 30%…
How Graph Neural Networks Are Accelerating Drug Discovery
In this episode of The Data Science Podcast, Lucas and Luna dive into how graph neural networks are transforming pharmaceutical research. They anchor on a real breakthrough: a team at MIT and Harvard used a GNN to screen over 100 million molecules and discover a new class of antibiotics effective against MRSA. The hosts explain why traditional drug discovery is slow and costly, and how treating molecules as graphs allows GNNs to learn complex molecular properties from limited data. They also…
How Data Scientists Build Natural Language Interfaces for Databases
In this episode, Lucas and Luna explore how data scientists are using large language models to let business users query databases with plain English. They dive into a real-world case from a hospital network that cut reporting time by 60% using a custom NL-to-SQL pipeline. The hosts discuss the architectural choices—smaller fine-tuned models vs. large APIs—and the critical guardrails needed to prevent hallucinated queries from corrupting production data. They also touch on cost management…
Showing the latest 50 episodes. The full archive of 183 is on Apple Podcasts, Spotify and every major podcast app — or via the RSS feed above.