Dr. Jan Philip Göpfert

Technical lead alongside founders from private equity and enterprise IT, building software that produces companies' CSRD sustainability reports. I scope what we build and then build it: an LLM pipeline from document intake to the finished report, in which every extracted claim traces back to its place in the source.

Built the evaluation we use to choose models and providers. It plants confounders, removes information and degrades the documents, and a reviewer can see exactly where on the page a value was found or missed.

The system runs on several model providers with failover, tracing on every call and data held in the EU.

Hands-on AI workshops for industry and academia.

Reviewed a healthcare startup's platform before launch (access control, multi-tenancy, observability), then built its LLM reporting pipeline, with every figure cited to its source.

Rewrote the training pipeline around a contrastive loss. It ran 200x faster than the one I found when I joined.

Built a multimodal semantic search engine on CLIP embeddings, from first prototype to production, and demoed it live to enterprise prospects. The demos won the customers.

Introduced CI/CD, Pydantic schemas and Parquet across the ML code.

Research on adversarial robustness, uncertainty quantification and human-machine interaction in Barbara Hammer's group.

Completed the doctoral thesis "Robustness in Machine Learning: Adversarial Perturbations, Explanations and Intuition" (2022).

Supervised 15+ Bachelor and Master theses.

Talks at international machine learning conferences; papers at ICANN, ESANN and IDA, and in Neurocomputing and IEEE Transactions on Human-Machine Systems.

Machine learning applied to 3D scanning and geometry processing, including 3D head reconstruction using CNNs.

Taught the Scientific Computing lecture.

Researched Monte Carlo methods and predictive analytics for the SAP HANA in-memory database.

Market analysis and prototypes for enterprise-scale analytics.