DANIEL SCHROTER THÜM

Prompt optimization for large language models is critical yet mostly a manual process. Engineers spend countless hours tweaking prompts without a systematic way to evaluate them against business metrics or their own data.

Promptic delivers automated, analytics-driven prompt optimization. Upload your data, provide the initial prompt, and let Promptic iteratively refine it using deep-learning models—no coding required. The output is a tailored prompt that maximizes downstream performance for your use case.

Data science consultants must stay up to date with the latest papers, yet newsletters rarely focus on the specific topics required for a project. It is difficult to track research across multiple fields in the fast-moving AI landscape.

PulsePicker is a customizable research newsletter. You subscribe to niche topics, an AI agent scans arXiv, and every week you receive a concise, tailored summary in your inbox.

CVs, one-pagers, and intranet profiles are hard to maintain and quickly become outdated.

Upload your documents, extract your skills automatically, and generate a polished one pager selling you as the expert you are.

Why are some arguments more convincing than others? Often it depends on the underlying human values they appeal to. We built the winning system of SemEval-2023 Task 4 to automatically detect those values in arguments.

Argument: “We should ban whaling, whaling is wiping out species for little in return.”

Predicted value: Universalism – Nature (confidence 0.92). The top three highlighted terms account for 78% of the model’s confidence. Orange terms support the prediction, showing that the model learned the relationship between “ban”, “wiping out species”, and the human value of protecting nature.

We participated in SemEval-2023 Task 4: identifying human values behind arguments. Our system topped 39 research teams and demonstrates that high-performing models can still be interpretable. We coupled strong architectures with explainable AI techniques, surveying end users to verify that the explanations are understandable.

“I’m working on a sentiment model with Python—who can help me?” Remote work and flexible models make it hard to know which colleagues hold certain skills.

Create direct, horizontal collaboration links. Let people discover expertise within the organization so that the workforce can respond to rapid change and reach its full potential.

NLP extracts competencies from the daily work. Employees receive a private profile to monitor their skills and can opt into being discoverable—forging new collaboration links across the organization.

Public debates increasingly happen online. Capturing key ideas from thousands of arguments can transform policy and business decisions. We built models that assign arguments to key points and, crucially, explain why the match occurs.

We modified LIME to work with siamese neural networks, allowing us to visualize why an argument matches a key point. Orange highlights boost similarity while blue highlights decrease it.

Argument: “The United States is undoubtedly the richest country ...”

Key point: “The US has a good economy / high standard of living.”

Prediction probabilities: Similar 0.78, Dissimilar 0.22. The explanation shows that terms such as “income” and “productive country” contribute to the model’s confidence, even if they are not literally part of the key point.

We address the KPA shared task 2021 of matching arguments to key points. Two architectures are explored: a TSDAE combined with a siamese network and a transformer-based siamese model enriched with linguistic features.

Both improve upon prior benchmarks by 2–5%. To make decisions explainable, we introduce a simple yet novel adaptation of LIME for siamese networks. The project contributes both stronger models and a practical interpretability technique for argument-key point matching.

I modelled the beer distribution game as a reinforcement learning problem to understand whether an agent can dampen the bullwhip effect. Agents representing different supply-chain participants (random, human-like, optimal) interact and the RL agent learns an ordering policy that reduces inventory costs by up to 69%.

Human decision makers frequently create costly fluctuations in serial supply chains. The beer distribution game illustrates these dynamics but the optimal policy is already known. In many cases, continuous versions of the game are more realistic, allowing us to leverage policy-gradient-based reinforcement learning.

I created both discrete and continuous environments to experiment with REINFORCE and DDPG algorithms. The results show that policy-gradient methods can match state-of-the-art action-value approaches while scaling better to multi-product settings. The work lays the foundation for future research on intelligent agents controlling supply chains.

This project was part of an internship with a financial institution, therefore I cannot share source code.

Build a Rasa-driven chatbot to support front-office teams. The bot was paired with a Haystack-based question-answering pipeline and Elasticsearch document store to fetch precise answers from policy documents. Feel free to reach out if you’d like to learn more.

Within this thesis I applied various machine learning techniques to solve a predictive maintenance task within a production line. Using the BOSCH competition dataset, I walked through the full CRISP-DM process and built a solution that has tangible business impact.

Feel free to download the PDF to dive deeper.

In the era of Industry 4.0, companies need practical showcases on how to apply AI to real problems. This case focuses on predicting whether a produced part fails internal quality control so that only high-quality goods are shipped.

The case balances business perspective with statistical rigor. It highlights how creativity, experimentation, and compromises are required across the entire data-mining process. It is designed as a teaching case for business students as well as decision makers dealing with predictive analytics.