Data Science Conversations

Steel has been shaped by fire and force for centuries. What happens when you add intelligence to that equation?
This episode goes inside the AI operation at SHS, a major German steel group, where a 20-person interdisciplinary team is embedding AI across production, research, and administration. Guests Michael Schaefer, Anna Volker, Ulrike Faltings, and Tobias Bettinger cover how they grew from three people in 2017, the challenges of legacy systems and messy industrial data, and why close collaboration with domain experts has been key to their success, with honest reflections on data quality, model monitoring, and the future of AI in steel.
Key Topics:
Modern Steel Plant Explained — Today's steel facilities operate as digital ecosystems of sensors, simulation, and AI. High-quality steel, used in automotive and offshore wind, demands far greater precision than commodity alternatives.
Building the AI Team — Starting with three people in 2017, SHS's AI function has grown to 20 specialists across three sub-teams: Specialised AI, Gen AI, and R&D — deliberately interdisciplinary across physics, maths, computer science, and engineering.
Infrastructure and Data Complexity — SHS runs its own data centres for real-time, low-latency production control. Managing a heterogeneous landscape of legacy systems and up to 10,000 sensors creates significant integration challenges.
Data Quality and Pipelines — Messy data — from Excel records and missing values to sensor drift — is the biggest obstacle to AI development. Robust pipelines depend on domain expert input, strict guardrails, and continuous model monitoring.
High-Impact Production Use Cases — Standout projects include an input material cost optimisation model, a defect detection system that saved a major customer relationship, and AI-driven temperature and oxygen prediction models that outperform traditional physical approaches.
Gen AI in Administration — LLMs are being used to automate feasibility analysis and process unstructured customer enquiries via a centralised, governed internal platform. Gen AI is kept out of live production due to hallucination risks, with human-in-the-loop built into all workflows.
Domain Expert Collaboration — Plant engineers and operators are central to every stage of AI development — shaping pipelines, detecting model failures, and bringing unsolved problems to the team. Years of shared successes have built deep mutual trust and a genuine two-way knowledge exchange.
Future Outlook and Advice — SHS aims to make the entire group AI-powered. Key advice for data scientists in heavy industry: understand processes and people before algorithms, embrace imperfect data, and play the long game. Smaller specialised models are an exciting near-term development; physical AI a longer-term frontier.

Michael Haft, founder of xplain Data, discusses causal discovery and causal AI, explaining how understanding cause-and-effect relationships goes beyond predictive modeling to enable truly intelligent interventions. He explores the technical foundations of object analytics, real-world applications in healthcare and manufacturing, and his vision for integrating causal AI into future intelligent systems.
Episode Summary
Causal Discovery vs. Prediction - Causal discovery aims to understand why things happen rather than just predicting what will happen. Unlike predictive models that rely on correlations, causal discovery identifies true cause-and-effect relationships necessary for intelligent interventions and goal achievement.
The Confounder Challenge - Understanding causality requires comprehensive data to identify confounders—hidden common causes that create spurious correlations. The gray hair and glasses example illustrates how age acts as a confounder, making the two correlated without a direct causal relationship between them.
Object Analytics Technology - Traditional machine learning requires flat tables, but real-world data (like electronic health records with 150+ tables) is inherently complex. Object analytics allows algorithms to work with comprehensive, holistic data structures, enabling deeper causal analysis without manual feature engineering.
Manufacturing Use Case - A cylinder head manufacturing example demonstrates how causal discovery identified the complete pathway from washing machine timing through part temperature to false negative leakage test results, enabling an intelligent process intervention that traditional predictive models couldn't provide.
Healthcare Applications - Projects using MIMIC hospital data analyze causes of pressure injuries in patients. The vision is to provide doctors with causal knowledge derived from millions of patient records to improve treatment decisions, discover new drug effects, and enable cost-efficient healthcare.
Path to Causal Maturity - Organizations need education on the difference between prediction and causality, comprehensive data availability, and engagement from both business owners (who have problems to solve) and data science teams. The shift requires iterative learning and hands-on experience with the technology.
Community Edition Launch - Explained Data is releasing a community edition starting with pre-configured object analytics models for the MIMIC healthcare dataset, followed by a full version for the broader data science community, with free access for universities and evaluation purposes.
Future of Causal AI - The next generation of AI systems will integrate causal layers with large language models, moving beyond text rephrasing to answering "why" questions based on empirical cause-and-effect relationships, particularly transforming healthcare and enabling more explainable, intelligent decision-making systems.

In this episode, we had the privilege of speaking with Akhil Patel, a globally recognized expert in economic cycles, discusses the 18-year boom-bust pattern and warns that we're approaching the peak of the current cycle in 2026, with a major financial crisis likely in 2027. He analyzes the AI investment bubble, draws parallels to historical manias, and provides practical strategies for businesses and investors to prepare for the downturn.
Episode Summary
1. Understanding Economic Cycles - Akhil Patel explains why cycles matter, emphasizing that cyclical patterns appear throughout nature and human behavior, particularly in stock markets and economies. Understanding these rhythms helps predict both prosperity and crisis periods.
2. The 18-Year Cycle Theory - the hypothesis of a regular 18-year boom-bust cycle (sometimes 16-20 years) in Western economies, particularly the US and UK. This pattern, first identified by economist Homer Hoyt in the 1930s through Chicago land sales data, has preceded every major financial crisis over the past 200 years.
3. Land Values Drive Cycles - Land is identified as the key indicator because it's a scarce, monopolistic asset that captures economic surplus. Property prices and speculation patterns serve as the primary mechanism driving both the boom and bust phases, with banking credit amplifying these movements.
4. Current Cycle (2011-2026) - Walking through the present cycle, Akhil identifies 2011-2012 as the starting point following the 2008 crisis. The COVID pandemic compressed what would normally be a 7-year second half into just 2 years of mania (2020-2022), though we're still seeing bubble behavior in AI investments arriving on schedule.
5. AI Investment Bubble Analysis - The current AI sector exhibits classic bubble characteristics: inflated valuations disconnected from fundamentals, enormous capital investment with questionable returns, and incestuous interconnections between major players (Nvidia, OpenAI, Oracle). Parallels are drawn to the dot-com bubble, 1980s Japan, and 19th-century railway booms.
6. Crisis Timing: 2026-2027 - Akhil predicts the property market will peak in 2026, with a major financial crisis following 6-12 months later in 2027. The trigger location is uncertain but likely in areas with extreme speculation—possibly the Middle East, parts of Asia, or unexpectedly in Germany, rather than the US which remains cautious after 2008.
7. Practical Preparation Strategies - Key recommendations include: avoid leverage, build cash reserves, ensure businesses can survive revenue declines, don't buy based solely on capital gains momentum, and position to acquire assets during the downturn. The advice emphasizes survival first, then opportunistic expansion during recovery.
8. Future Outlook Beyond Crisis - Despite the predicted downturn, Akhil remains optimistic about the next cycle (post-2030), believing AI and blockchain technologies are genuinely transformative once properly applied. The tech sector typically leads recovery, offering significant opportunities for those who survive the crisis with resources intact.

Peter Bärnreuther from Munich RE discusses the emerging field of AI insurance, explaining how companies can manage the inherent risks of probabilistic AI systems through specialized insurance products. The conversation covers real-world AI failures, different types of AI risks, and how insurance can help both corporations and AI vendors scale their operations safely.
Key Topics Discussed
Peter's Career Journey: Peter Bärnreuther transitioned from studying physics and economics to risk management at Accenture, then Munich RE, where he developed crypto insurance products before joining the AI risk team to create coverage for AI-related risks.
Probabilistic vs Deterministic Systems: Unlike traditional deterministic systems where errors can be traced, AI systems are probabilistic - they can be 99.5% accurate but never 100% certain, creating fundamental new risks that require insurance coverage.
AI Risk Categories: Two main types exist - traditional machine learning risks (classification errors like fraud detection) and generative AI risks (IP infringement, hallucinations, legal compliance issues), each requiring different insurance approaches.
Real-World AI Incidents: Examples include airline chatbots promising unauthorized discounts, lawyers using fake legal cases, and AI house valuation systems losing $300M+ by failing to adjust to market changes during price drops.
Insurance Product Structure: Munich RE offers two main products - one for corporations using AI internally for risk mitigation, and another for AI vendors needing trust-building to scale their business and attract enterprise clients.
Specific Use Cases: Successful implementations include solar panel fault detection (100% accuracy guarantee), credit card fraud prevention (99.9% performance guarantee), and battery health assessment for electric vehicles with compensation guarantees.
Market Challenges: Key difficulties include pricing models with limited historical data, concept drift where AI performance degrades over time, accumulation risk when multiple clients use similar foundation models, and "silent coverage" issues in existing insurance policies.
Future Market Outlook: AI insurance may either become a separate line of business (like cyber insurance) or be integrated into traditional policies, with current focus on US and European markets and strongest traction in IT security applications.

Chris Parmer (Chief Product Officer & Co-Founder, Plotly) and Domenic Ravita (VP of Marketing, Plotly) discuss the evolution of AI-powered data analytics and how natural language interfaces are democratizing advanced analytics.
Key Topics Discussed
AI's Market Category Convergence Domenic describes how AI is collapsing traditional boundaries between business intelligence tools (Power BI, Tableau), data science platforms, and AI coding tools, creating a quantum leap similar to the drag-and-drop revolution 20 years ago.
The 30/70 Engineering Reality Chris reveals that LLMs represent only 30% of AI analytics products, with 70% being sophisticated tooling, error correction loops, and multi-agent systems. Raw LLM output succeeds only one-third of the time without extensive supporting infrastructure.
Code-First AI Architecture Plotly's approach generates Python code rather than having AI directly process data, creating more rigorous analytics. The system generates 2,000-5,000 lines of code in under two minutes through parallel processing while maintaining 90%+ accuracy.
Natural Language as Universal Equalizer Discussion of how natural language interfaces eliminate the learning curves of different analytics tools (Salesforce, Tableau, Google Analytics), potentially democratizing data visualization across organizations by providing a common interface.
Vibe Analysis Concept Introduction of "vibe analysis" - the data equivalent of "vibe coding" - enabling fluid, rapid data exploration that keeps analysts in flow states through natural language interactions with AI-powered tools.
Transparency and Trust Building Exploration of building user trust through auto-generated specifications in natural language, transparent logging interfaces, and making underlying code assumptions visible and adjustable to prevent misleading results.
Human-AI Collaboration Balance Chris emphasizes that while AI accelerates visualization creation and data exploration, human interpretation remains essential for generating insights. The risk lies in systems that attempt to "skip to the finish" with fully automated decision-making.
Infrastructure Misconceptions Domenic predicts people will wrongly assume AI analytics requires extensive data warehouses and semantic layers, when effective analysis can work with standard databases and file formats, making advanced analytics more accessible than many realize.