Projects

Open-source framework that drives multi-turn LLM dialogue through hierarchically nested state machines, giving developers more deterministic control where free-form prompting is too brittle for the task.

• Co-developed with ZHAW and released publicly under the MIT license.
• Published in CAiSE 2024 proceedings.
• Underpins the dialogue orchestration in our digital health and psychotherapy agents.

Role

Co-developer and co-author

Partner

ZHAW

AI-augmented persuasive system that helps students act on career advice instead of postponing it, closing the loop between a counselling session and the steps a student is supposed to take afterwards.

• Targets procrastination as the point where career guidance breaks down.
• Pre-post evaluation with 15 students showed lower state procrastination and higher career decision-making self-efficacy, with most counselling-defined goals met or exceeded.
• Manuscript in preparation with collaborators at UZH.

Role

Design, evaluation, and technical advising

Domain

University career services

Method

Action design research

Evaluation

Mixed-methods pre-post, 15 students

Advising public-sector teams on LLM use cases, system design, and privacy-compliant deployment, alongside empirical work on how citizens and public servants judge the legitimacy of AI agents that represent citizen voices in the policy cycle.

• Co-authored a published study on AI agents acting as digital representatives for citizens in municipal policy formulation.
• Exploratory interviews with an embedded vignette gathered views from 52 citizens and public servants in a German municipality.
• Both groups saw AI representation as potentially legitimate, under conditions that could move participation from consultation toward a structural role in governance.

Role

Technical advisor

Conversational agent that maintains continuity between medical consultations through personalized dietary and activity messaging, keeping patients engaged with their care plan in the weeks when no clinician is in the room.

• Field-tested with 29 users in clinics across Canton Zurich.
• Persuasive prompting strategies derived from the design work were published at ECIS 2025.
• A follow-up paper under review at Computers in Human Behavior examines how patients perceive chatbot empathy and what makes it persuasive.

Role

Design, development, evaluation

Funding

Innosuisse

Partner

ZHAW, UZH, COBEDIAS, Helmedica AG

Master's thesis project projecting text zoning labels across languages so that job advertisements in a low-resource language can be segmented without hand-annotated training data.

Role

Author

Institution

University of Zurich