A Master thesis topic is available at the Centre for Digital Health Interventions!

Overview

We supervise a small number of Master’s theses each year on the STELLA platform. Projects are scoped so that a well-executed thesis results in a peer-reviewed submission and a merged contribution to an open-source system in active use.

This is an open call rather than a single predefined topic. We are primarily interested in students who bring a research question of their own within the problem space described below. If you do not have one yet, we can match you to a scoped topic.

Strong software engineering ability is a prerequisite. Thesis work is integrated into a platform that other research groups depend on and that supports an upcoming clinical study.

Project Purpose and Vision

For most users, voice is one interface among several. For an older adult with early-stage dementia, it is frequently the only accessible one. This makes voice agents unusually valuable for this population, and unusually consequential when they fail.

Existing systems force an unsatisfactory trade-off. Commercial voice platforms achieve low latency but operate as closed architectures with no inspectable reasoning. Open-source frameworks provide transparent audio infrastructure but leave safety arbitration, session management, and conversational structure unaddressed. Neither is adequate for clinical research.

We develop STELLA, an open-source platform for auditable, low-latency, voice-first conversational agents, and GRACE, an agent delivering Cognitive Stimulation Therapy to individuals with early-stage dementia. GRACE is not one application among many. It functions as the design specification. A user who pauses to retrieve a word, who trusts the agent’s output, and who has no means of verifying it, generates every requirement that matters also for a broader audience: turn-taking that does not misread a retrieval pause as a completed turn, principled restraint about when not to intervene, and an audit trail sufficient for regulatory inspection. Designing for this user derives the rest of the platform’s requirements.

Project Goals

A thesis in this group advances one of the following areas. The strongest projects connect two.

• Safety under adversarial conditions. Voice-channel prompt injection, jailbreak resistance, and the evaluation of conversational risk in the absence of ground truth and under non-deterministic system behaviour.
• The interaction layer. Endpointing that distinguishes word-finding pauses from turn completion, expressive speech synthesis appropriate to a clinical register, and voice characteristics that sustain trust across repeated sessions.
• Affect sensing and adaptation. Emotion and stress detection from speech, video, or both. The substantive question is not detection accuracy but when acting on the signal supports the user rather than manipulating them.
• Dialogue policy. Probing depth, persistence, and graceful abandonment. Over-questioning a vulnerable user constitutes a clinical harm rather than a usability defect.
• Platform infrastructure. Plan authoring, state management, evaluation tooling, and developer experience.

Thesis Structure

The student defines a research question, either their own or one drawn from our scoped topic list. They implement the contribution within STELLA and integrate it into the running platform. They then design and conduct an empirical evaluation and prepare the results to publishable standard.

Required Skills

• Ability to work independently in an existing production codebase
• Comfort with a Scrum process, code review, and shipping into a codebase others depend on
• Proficiency in Python or TypeScript, ideally both
• Substantive experience building with LLMs beyond introductory API use
• Familiarity with experimental design and empirical evaluation
• German language skills are an advantage but not a requirement

Supervision