Episodes

  • How Data Scientists Build Guardrails for Large Language Models

    How Data Scientists Build Guardrails for Large Language Models
    Jul 14 2026
    Episode 109 of The Data Science Podcast explores how data scientists are building guardrails to keep large language models safe, accurate, and on-brand in production. Lucas and Luna walk through a real case: a fintech chatbot that hallucinated a fake regulatory filing. They break down the guardrails stack — input validation, output moderation, and continuous monitoring — using concrete examples like NVIDIA's NeMo Guardrails and open-source tools like Guardrails AI. They also discuss the tension between user experience and safety, and why guardrails are the new CI/CD for LLM ops. If you're deploying generative AI, this episode gives you a practical framework for catching failures before they reach users. #LLMGuardrails #AISafety #GenerativeAI #DataScience #MachineLearning #NVIDIANeMo #GuardrailsAI #LLMOps #AIHallucination #PromptInjection #Fintech #Chatbot #ModelGovernance #AIDetection #MLProduction #Technology #FexingoBusiness #BusinessPodcast Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    11 mins
  • How Data Scientists Are Building AI Agents That Actually Work

    How Data Scientists Are Building AI Agents That Actually Work
    Jul 13 2026
    Lucas and Luna dive into the practical reality of AI agents in mid-2026 — not the hype, but the actual engineering choices that make them reliable. They unpack a concrete case: a mid-size logistics company that deployed a multi-agent system to handle shipment rerouting during the 2025 hurricane season. Lucas walks through the agent architecture — a coordinator agent, a weather data agent, a routing agent, and a customer comms agent — and explains why the team chose a deterministic fallback layer over pure LLM autonomy. Luna challenges whether agents are just chatbots with extra steps and pushes Lucas on where the data science value really lives. The episode covers agent orchestration frameworks (LangGraph vs. custom state machines), the role of synthetic data for testing edge cases, and why retrieval-augmented generation is the unsung backbone of production agents. Listeners walk away with one concrete pattern: the supervisor agent pattern with human-in-the-loop for high-stakes decisions, and a clear sense of what separates a demo from a deployment. #AI_Agents #MultiAgentSystems #LLM #AgenticWorkflow #LangGraph #Orchestration #RetrievalAugmentedGeneration #ProductionML #DataScience #Logistics #WeatherData #SyntheticData #HumanInTheLoop #SupervisorAgent #MachineLearning #Technology #FexingoBusiness #BusinessPodcast Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    8 mins
  • How Data Scientists Use Data Version Control for Reproducibility

    How Data Scientists Use Data Version Control for Reproducibility
    Jul 13 2026
    Lucas and Luna break down why data version control (DVC) has become as essential as Git for machine learning teams. They trace the problem through a concrete example: a fraud detection model at a fintech company where a missing dataset version caused a 15 percent drop in recall. The episode walks through how DVC tracks data snapshots, pipeline stages, and model artifacts—without duplicating massive files—using a simple declarative YAML config. Lucas explains the difference between DVC's approach and Git LFS, and why tools like Pachyderm and DVC solve overlapping but distinct problems. The hosts also discuss how versioning interacts with feature stores and CI/CD for ML, and why the field is moving toward treating data with the same discipline as source code. No fluff, just a focused look at one practice that separates professional data teams from the rest. #DataVersionControl #DVC #MLOps #Reproducibility #MachineLearning #DataScience #GitForData #Pachyderm #LFS #DataPipeline #FeatureStore #CI/CD #FraudDetection #Fintech #MLPipeline #DataGovernance #Technology #FexingoBusiness Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    13 mins
  • How Data Scientists Use Feature Stores for Reproducible ML

    How Data Scientists Use Feature Stores for Reproducible ML
    Jul 12 2026
    In episode 106 of The Data Science Podcast, Lucas and Luna dive into the practical world of feature stores—centralized repositories for machine learning features. They explore how companies like Uber and Netflix use feature stores to ensure reproducibility, reduce duplication, and speed up model deployment. Lucas breaks down the architecture of a typical feature store, including offline and online serving, while Luna shares a real-world example from a fintech startup that cut feature engineering time by 60 percent. They also discuss the trade-offs between open-source solutions like Feast and managed offerings from cloud providers. By the end, you'll understand why feature stores are becoming a critical part of the MLOps stack. #FeatureStore #MLOps #DataScience #Reproducibility #FeatureEngineering #Uber #Netflix #Feast #MachineLearning #DataEngineering #OfflineStore #OnlineStore #FeatureServing #DataPipeline #ModelDeployment #Technology #FexingoBusiness #BusinessPodcast Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    10 mins
  • How Data Scientists Use Federated Learning for Privacy-Preserving ML

    How Data Scientists Use Federated Learning for Privacy-Preserving ML
    Jul 12 2026
    Episode 105 dives into federated learning, the privacy-preserving technique that trains models across decentralized data without ever centralizing sensitive information. Lucas and Luna unpack a real-world case: how Apple uses federated learning to improve QuickType keyboard predictions on iPhones without sending your typing data to the cloud. They break down the key technical components — local model training, secure aggregation, and differential privacy — and explain the trade-offs: communication cost vs. accuracy, and the challenge of non-IID data across thousands of devices. The conversation also touches on Google's Gboard implementation and how healthcare researchers are exploring federated learning for multi-hospital models without sharing patient records. Listeners will walk away understanding both the mechanics and the real-world constraints of one of the most important privacy technologies in modern machine learning. #FederatedLearning #PrivacyPreservingML #Apple #QuickType #Gboard #Google #SecureAggregation #DifferentialPrivacy #EdgeComputing #HealthcareAI #DataPrivacy #MachineLearning #Tech #FexingoBusiness #BusinessPodcast #Technology #DataScience #AI Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    11 mins
  • How Data Scientists Use Gradient Boosting for Tabular Data

    How Data Scientists Use Gradient Boosting for Tabular Data
    Jul 11 2026
    A deep dive into the enduring power of gradient boosting machines (GBMs) for structured, tabular data—the bread and butter of most real-world data science. Lucas and Luna explore why gradient boosting consistently wins Kaggle competitions and beats deep learning on many business problems. They break down the core mechanics: sequential tree-building, learning rate, and regularization. The episode focuses on a case study from a mid-size e-commerce company that used XGBoost to reduce customer churn prediction error by 18% year-over-year. They also discuss modern variants like LightGBM and CatBoost, and when to choose each. Practical guidance on hyperparameter tuning and common pitfalls (overfitting, categorical encoding) grounds the conversation in daily data-science work. Listeners will walk away understanding why gradient boosting remains a must-have in any data scientist's toolkit, especially for data with mixed data types and missing values. #GradientBoosting #XGBoost #LightGBM #CatBoost #TabularData #MachineLearning #DataScience #Kaggle #HyperparameterTuning #ChurnPrediction #EnsembleMethods #DecisionTrees #Regularization #FeatureEngineering #BusinessAnalytics #Technology #FexingoBusiness #BusinessPodcast Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    9 mins
  • How Data Scientists Use Monte Carlo Simulations for Risk

    How Data Scientists Use Monte Carlo Simulations for Risk
    Jul 11 2026
    Episode 103 of The Data Science Podcast with Fexingo. Lucas and Luna dive into Monte Carlo simulations — not as a textbook concept, but as a practical tool data scientists use to quantify uncertainty. They walk through a real-world case: a mid-size logistics company that used Monte Carlo to model delivery times under variable traffic, weather, and fuel costs. Lucas explains the math behind random sampling, how to choose the number of simulations, and the common pitfall of assuming normal distributions. Luna challenges him on interpretability — how do you explain a distribution of outcomes to a non-technical stakeholder? They also discuss modern libraries like NumPy and PyMC, and how cloud computing has made millions of simulations feasible on a laptop. No abstract theory — just a grounded look at when Monte Carlo beats deterministic models and when it doesn't. By the end, you'll know exactly how to frame a Monte Carlo problem for your next data science project. #MonteCarlo #RiskSimulation #DataScience #UncertaintyQuantification #NumPy #PyMC #Logistics #PredictiveModeling #Simulation #BusinessAnalytics #MachineLearning #Probability #DecisionMaking #StochasticModeling #Technology #FexingoBusiness #BusinessPodcast #DataDriven Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    9 mins
  • How Data Scientists Use SBERT for Semantic Search at Scale

    How Data Scientists Use SBERT for Semantic Search at Scale
    Jul 10 2026
    In this episode, Lucas and Luna dive into the practical applications of Sentence-BERT (SBERT) for semantic search in production. They discuss how SBERT converts text into dense vector embeddings, enabling similarity search beyond keyword matching. The hosts walk through a real-world case study of a mid-sized e-commerce company that replaced its legacy Elasticsearch-based search with an SBERT-powered semantic search, reducing the number of searches that return zero results by 40 percent, and cutting the cost of maintaining a custom synonym list by $100,000 annually. They also cover trade-offs: the need for GPU infrastructure during embedding generation, the latency vs. accuracy balance using approximate nearest neighbor algorithms, and how fine-tuning on domain-specific data improved relevance by 15 percent. The episode closes with a reflection on when to use SBERT versus newer large language models for search. #DataScience #SemanticSearch #SBERT #SentenceBERT #NLP #VectorEmbeddings #ApproximateNearestNeighbors #Elasticsearch #Ecommerce #MachineLearning #Technology #SearchEngines #FineTuning #BERT #Embeddings #ProductionML #FexingoBusiness #BusinessPodcast Keep every episode free: buymeacoffee.com/fexingo
    Show More Show Less
    9 mins