In this episode, Lucas and Luna dive into multimodal zero-shot learning, the technique that lets AI models like CLIP recognize objects, scenes, and text across images without ever being explicitly trained on those combinations. They explore a concrete use case: a retail startup using a pretrained multimodal embedding model to automatically tag 10,000 new product photos per day with zero labeling cost. Lucas breaks down the architecture—contrastive learning on image-text pairs—and explains why zero-shot works when the embedding space is aligned. Luna asks about failure modes: ambiguous images, domain shift, adversarial inputs. They also touch on the trade-off between generality and fine-tuning, and where the field is heading next. No hype, just how the math makes it possible.

#MultimodalLearning #ZeroShotLearning #CLIP #ContrastiveLearning #DataScience #MachineLearning #ImageRecognition #NaturalLanguageProcessing #Embeddings #AI #RetailTech #ProductTagging #TransferLearning #DeepLearning #ComputerVision #Technology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

In Episode 95 of The Data Science Podcast with Fexingo, Lucas and Luna dive into a practical yet underappreciated technique: using k-nearest neighbors for anomaly detection. They kick off with a real-world story from a major credit card processor that flagged a series of fraudulent transactions by measuring distance to the nearest legitimate patterns. Lucas explains why distance-based methods can outperform deep learning in low-signal, high-stakes settings, especially when you need interpretable reasons for each flag. Luna challenges him on scalability and the curse of dimensionality, and they discuss how companies like Stripe and PayPal have used variants of k-NN in production fraud pipelines. They also touch on the trade-offs between global and local outlier factors, and how to choose k when the definition of 'normal' shifts over time. A concrete segment on choosing distance metrics — Euclidean vs. Manhattan vs. cosine — gives listeners an actionable guideline. Mid-episode, they weave in a natural request for listener support, tying it back to the value of open-source tools. If you've ever wondered when to reach for a simple nearest-neighbor approach instead of a neural network, this episode gives you the framework.

#DataScience #MachineLearning #AnomalyDetection #KNearestNeighbors #FraudDetection #OutlierDetection #DistanceMetrics #LocalOutlierFactor #Stripe #PayPal #CurseOfDimensionality #Interpretability #Technology #FexingoBusiness #BusinessPodcast #TechPodcast #DataSciencePodcast #ProductionML

Keep every episode free: buymeacoffee.com/fexingo

In episode 94 of The Data Science Podcast with Fexingo, Lucas and Luna explore how active learning cuts labeling costs by 80 percent while maintaining model accuracy. Using a concrete example from a medical imaging startup training a rare-disease classifier, they walk through uncertainty sampling, query strategies, and the human-in-the-loop workflow. They compare pool-based versus stream-based active learning, discuss common pitfalls like distribution shift, and explain when active learning beats random sampling. If you are a data scientist looking to stretch a limited labeling budget, this episode gives you a practical framework to get started. Lucas and Luna also touch on tools like modAL, scikit-activeml, and Label Studio. No hype, just signal.

#ActiveLearning #DataScience #MachineLearning #Labeling #UncertaintySampling #HumanInTheLoop #MedicalImaging #RareDisease #modAL #scikitActivelm #LabelStudio #DistributionShift #QueryStrategy #PoolBased #StreamBased #Technology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

Episode 93 of The Data Science Podcast with Fexingo dives into Thompson Sampling, a Bayesian approach to online experimentation that balances exploration and exploitation better than traditional A/B testing. Lucas and Luna walk through a concrete example from a real e-commerce site that ran a 50-variant landing page test — and how Thompson Sampling found the winner in half the time with 30% less traffic wasted. They also discuss Thompson Sampling's role in multi-armed bandit problems, how it handles changing user behavior, and why it's becoming a go-to technique for data scientists in marketing and product optimization. Plus: a quick note on how listener support keeps the podcast ad-free.

#ThompsonSampling #BayesianInference #MultiArmedBandit #OnlineExperiments #ABTesting #ExplorationExploitation #DataScience #MachineLearning #MarketingOptimization #ProductExperimentation #ECommerce #ConversionRate #BetaDistribution #BayesianStatistics #SequentialTesting #DataDriven #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

Episode 92 of The Data Science Podcast with Fexingo dives into embedded analytics—the practice of integrating dashboards, reports, and AI insights directly into customer-facing products. Lucas and Luna explore how companies like Notion and Canva use embedded analytics to drive product-led growth, reduce churn, and monetize data. They walk through a real case: a fictional B2B SaaS platform that cut customer onboarding time by 40% by embedding usage analytics inside its app. The hosts discuss the technical stack (frontend SDKs, API-first BI tools, semantic layers), the UX pitfalls to avoid (like overwhelming users with charts), and the business model shift from selling reports to embedding them. If you're a data scientist or product manager wondering how to turn dashboards into a growth engine, this episode gives you a concrete playbook.

#EmbeddedAnalytics #ProductLedGrowth #DataScience #BusinessIntelligence #Analytics #Notion #Canva #B2BSaaS #CustomerRetention #Monetization #SemanticLayer #APIFirst #UX #ProductAnalytics #Technology #FexingoBusiness #BusinessPodcast #DataDriven

Keep every episode free: buymeacoffee.com/fexingo

Most marketing attribution models are correlational — they tell you what happened, not why. In this episode, Lucas and Luna break down how data scientists are using causal inference techniques, specifically double machine learning and instrumental variables, to measure the true incremental impact of ad spend. Using a real 2025 case from a mid-market e-commerce brand that ran geo-lift tests across 50 DMAs, they show how naive last-click attribution overestimated Facebook ROI by 60 percent while underestimating podcast ads by 40 percent. The hosts explain why off-the-shelf attribution is broken, how double ML handles high-dimensional confounders like seasonality and competitor activity, and why the field is shifting from 'more data' to 'better questions.' Specific metrics, concrete numbers, no vague theory.

#CausalInference #MarketingAttribution #DataScience #DoubleMachineLearning #InstrumentalVariables #GeoLift #IncrementalMeasurement #ROI #DigitalMarketing #Tech #BusinessPodcast #DataDriven #Analytics #Econometrics #CausalEffect #Confounders #AdTech #FexingoBusiness

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna explore how knowledge graphs are supercharging retrieval-augmented generation (RAG) systems. They break down a concrete example: how a financial services firm used a knowledge graph built from SEC filings and earnings call transcripts to reduce hallucination in their Q&A chatbot by 40 percent. The hosts explain why flat vector search alone often fails, how graph traversal adds context, and what it takes to maintain a dynamic knowledge graph. They also touch on trade-offs like latency and engineering complexity. If you've wondered when RAG needs more than a vector database, this episode gives you the practical answer.

#KnowledgeGraphs #RAG #RetrievalAugmentedGeneration #Hallucination #VectorSearch #GraphTraversal #Neo4j #SECFilings #EarningsCalls #NLP #LLMOps #DataScience #Technology #FexingoBusiness #BusinessPodcast #MachineLearning #GraphDB #EntityResolution

Keep every episode free: buymeacoffee.com/fexingo

Lucas and Luna explore how graph neural networks are transforming recommendation systems, using the example of Pinterest's PinSage model. They break down how GNNs capture relational data like user-item interactions to generate high-quality recommendations, discuss the challenges of scaling to billions of nodes, and compare GNN-based approaches to traditional collaborative filtering. The episode includes a concrete explanation of message-passing in graphs and real-world performance metrics from Pinterest's deployment.

#GraphNeuralNetworks #RecommendationSystems #Pinterest #PinSage #MachineLearning #DataScience #Technology #CollaborativeFiltering #MessagePassing #NodeEmbeddings #GraphConvolution #Scaling #UserItemGraph #IndustrialML #Personalization #FexingoBusiness #BusinessPodcast #DataDriven

Keep every episode free: buymeacoffee.com/fexingo

Episode 88 of The Data Science Podcast dives into dimensionality reduction — but not for preprocessing. Lucas and Luna explore how data scientists use t-SNE and UMAP to visualize high-dimensional data, from customer segmentation to single-cell genomics. They discuss the trade-offs between global and local structure preservation, the risk of over-interpreting clusters, and why a 2D plot is never the whole truth. With concrete examples from retail analytics and biology, this episode gives you a practical framework for when to use t-SNE versus UMAP and how to avoid common pitfalls. If you've ever stared at a scatter plot and wondered if the patterns are real, this one's for you.

#DimensionalityReduction #tSNE #UMAP #DataVisualization #MachineLearning #DataScience #Clustering #HighDimensionalData #SingleCellGenomics #CustomerSegmentation #PCA #Interpretability #Visual Analytics #FeatureEngineering #UnsupervisedLearning #Technology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

In episode 87 of The Data Science Podcast, Lucas and Luna explore manifold learning—a powerful technique for dimensionality reduction that goes beyond PCA. They focus on t-SNE and UMAP, explaining how these algorithms preserve local and global structure in high-dimensional data. Using concrete examples from genomics and image datasets, they discuss when to use each method and common pitfalls like misleading visualizations. Lucas shares a cautionary tale about a team that over-interpreted a t-SNE plot, while Luna explains how UMAP scales to millions of points. They also touch on recent developments like parametric UMAP and its integration with deep learning. Perfect for data scientists who want to understand the trade-offs between linear and nonlinear dimensionality reduction.

#ManifoldLearning #DimensionalityReduction #tSNE #UMAP #DataScience #MachineLearning #PCA #DataVisualization #TopologicalDataAnalysis #Genomics #ImageData #KLDivergence #CrossEntropy #ParametricUMAP #FexingoBusiness #BusinessPodcast #Technology #DataSciencePodcast

Keep every episode free: buymeacoffee.com/fexingo