Join Us

● Software Engineering for Data Science: making software in Python that is useful for the lab and for
other labs. Either improving our existing software or developing new ideas that you can bring.
● Current main Software Engineering projects that are worth exploring further:
○ mmore, a processing and RAG pipeline used notably for preparing the training mixture of
Apertus, the Swiss LLM,
○ MMIRAGE, an LLM-driven data processing pipeline
● LiGHT Bootcamp: a MOOC with human supervision meant for medical doctors to learn about how AI
projects work, and for computer scientists to learn about the specificities of making AI with medical
data. This MOOC requires testing and improving.
● Meditron beyond English: MultiMeditron (turning Meditron into a multimodal model that takes medical
images and not just text) and Polyglot Meditron (training Meditron on low-resource languages and
evaluating its capabilities)

● Multilingual Clinical Speech Recognition
● Multilingual Medical LLM
● Model architectures for humanitarian decision support

● PRISM — Program for Research in Social Art and Media
Exploring AI, media, design, and social impact through creative and research-driven projects.
● LiGHT Innovation Hub
Developing a platform for innovation theory, rapid prototyping, and AI development for health and
humanitarian applications.
● LiGHT Ultratron
Building a multimodal agentic foundation model for ultrasound, including image understanding,
clinical reasoning, reporting, and workflow support.
● Mam-AI
Developing AI-supported tools for maternal health, obstetric ultrasound, and task-shifted care in
resource-limited settings.
● Exploratory project areas
Students may also work on smaller exploratory projects related to AI-enabled clinical workflows,
human-centered AI design, model evaluation, data engineering, deployment, or responsible innovation
frameworks.

● Open-ended clinical evaluation
● Synthetic data for medical LLMs
● Safety of medical LLMs
● Medical LLM training at scale (needs to be a master thesis / internship)

Student-proposed projects in this space also welcome.

● Structured reasoning paradigms & agentic frameworks for reliable Aİ
● Post-training for reasoning and clinical alignment
● Test-time scaling & inference-time reasoning for medical decision-making

Position paper - participation as an epistemic condition for health AI evidence The argument is that
community participation isn't just ethically desirable it's a validity requirement for post-market evidence in
health AI. Without it, what gets called "real-world evidence" is often just a technical audit of a system that
communities never consented to be evaluated by. The paper develops this through the Participatory Evidence
Threshold framework and situates it within existing regulatory thinking on SaMD and post-market
surveillance.
Living literature review - participatory evaluation in health AI governance A continuously updated review
that maps how participatory feedback loops have been theorised and (occasionally) implemented in clinical
AI regulatory cycles, particularly in LMIC settings. It feeds directly into the MOOVE framework and helps
surface where the literature is genuinely thin versus where it just hasn't been connected.
Playbook - participatory evaluation as health AI governance infrastructure A practical,
practitioner-facing document developed with ICMR-NIRDHDS, Ashoka University, and PrAImaan. The idea is
to translate the theoretical arguments around participatory evaluation into something that health system
actors — regulators, deployers, civil society organisations — can actually use. It covers how to design
evaluation processes that produce credible evidence, not just procedural legitimacy.