Applied ML Projects

Selected applied machine-learning projects conducted at HSLU.

Regulatory Information System 

LLM-based Information Extraction from Regulatory Documents

• Built an end-to-end OCR + LLM information-extraction system for ~1,500 heterogeneous legal and regulatory documents, transforming unstructured documents into structured, machine-readable data.

• Designed the extraction architecture around staged LLM reasoning and validation, separating relevance detection, value extraction, and consistency checking to handle noisy documents and ambiguous regulatory language.

• Developed a systematic evaluation framework for extraction quality and failure analysis, identifying error sources across OCR, retrieval/context selection, and LLM reasoning and using them to iteratively improve the pipeline.

ENFLATE

Machine Learning for Energy Forecasting

• Built an end-to-end machine-learning system for next-day energy/flexibility forecasting across heterogeneous real-world installations, from raw meter data and feature engineering to model evaluation and production inference.

• Developed and compared forecasting approaches ranging from robust statistical baselines and autoregressive models to nonlinear, pooled, and installation-specific methods, with particular focus on intermittency, regime changes, outliers, and cross-installation structure.

• Designed a leakage-free walk-forward evaluation and data-audit framework to uncover failure modes in both forecasting and target construction, quantify performance across installations, and ensure consistency between offline experiments and deployed forecasts.

Interpretable Decision-Making with Active Inference

Active Inference-based Community Management of Patient Communities in Healthcare (ongoing research project)

• Working on probabilistic models for interpretable, uncertainty-aware sequential decision-making using the Active Inference framework.

• Developing and evaluating computational approaches for adaptive community management, with emphasis on transparency and decision quality.