The Fifth Elephant - Pune Edition

Describe your session in 2 paragraphs

Digital twins, predictive maintenance models, and AI-driven quality control all promise to tell manufacturers what to do. But there’s a critical gap - most ML models only reveal correlations, not causation. When your predictive maintenance system flags high vibration as a failure predictor, does reducing vibration actually prevent failures, or is vibration merely a symptom of the real root cause like bearing wear or misalignment? When your quality model correlates temperature with defect rates, will adjusting temperature fix the problem, or is temperature confounded by shift changes and raw material batches? This distinction between “what predicts” and “what to intervene on” is the difference between optimizing your factory and chasing expensive red herrings.

This talk introduces Pearl’s ladder of causation as a practical framework for interrogating your Industrie 4.0 analytics—whether that’s a full digital twin or sensor data feeding a maintenance dashboard. Using a predictive maintenance scenario, I’ll demonstrate how causal inference tools—DAGs, the do-calculus, and adjustment sets—distinguish spurious correlations from actionable interventions. I’ll show why LLMs and standard ML are “causal parrots” that confuse P(Y|X) with P(Y|do(X)), and how to augment them with proper causal machinery. The result: explicit, auditable assumptions that let operations, engineering, and data science teams align on why an intervention should work before committing resources.

Mention 1-2 takeaways from your session

1. A clear framework (Pearl’s ladder) for recognizing when manufacturing decisions require causal inference rather than correlation-based ML—especially for root cause analysis, intervention planning, and counterfactual “what-if” scenarios
2. A practical workflow to encode domain knowledge as testable causal graphs (DAGs) and derive statistically valid adjustment strategies for predictive maintenance and quality control problems

Which audiences is your session going to be beneficial for?

Data scientists, analytics managers, and manufacturing/operations engineers building predictive maintenance systems, quality control models, or digital twins who want to move beyond correlation-based dashboards to causal, interventional decision-making. Also valuable for managers navigating disagreements between AI recommendations and shop-floor intuition.

Bio: I am a lapsed physicist with several years of experience in various Data Science and BI contexts, mostly in Germany. Now I am the founder of Romulan AI - building the causal layer for AI based digital transformation.