Lighthouse (Snackathon; Swiss-AI Lausanne)

We turn overwhelming diagnoses into clear, trusted guidance — giving families and therapists evidence-based next steps, therapies, and local support to bridge the gap from hospital to home.

Comment

Try the demo

Team Members

  • Iban Eguia
  • Mohamed Aziz Hamza
  • Javier Belmonte
  • África Santos
  • Deep Shah

Inspiration

When you receive a chronic diagnosis for your child, like cerebral palsy, it's a shock that transforms everything. But then reality hits: you must navigate complex medical information, the diagnosis, understand treatment flows, multiple specialists, and implement intensive rehabilitation plans that require daily execution.

The burden is crushing: fragmented information across dozens of websites, conflicting advice from different specialists, expensive equipment that may not work together, and the constant fear of missing something crucial. Families become project managers, researchers, and therapists overnight, while therapists struggle to provide consistent, evidence-based guidance across diverse cases and limited time.

This isn't just a medical problem—it's a human crisis that affects millions of families who need trustworthy, actionable guidance they can actually implement in their daily lives.

What it does

Our project bridges the gap between hospital diagnosis and daily care. We provide a trusted AI companion that translates medical knowledge into plain-language guidance, highlights evidence-based therapies, and points to local services and support networks.

Unlike generic AI chatbots, our model is built on curated, verified content, making it safe, reproducible, and auditable on hospital or public infrastructure.

Who benefits? Families, therapists, and caregivers who need clear, practical next steps.

Our core value: turning a diagnosis into an actionable care plan — from plain-language explanations, to summarized evidence, to therapy tools, devices, and local support.

Business model Hospitals, clinics and even insurances would be able to offer a tool to their customers to improve their day-to-day in living with these diseases and fighting misinformation. It would be a b2b business with health institutions.

How we built it

We have used an Open-WebUI interface hosted in HuggingFace spaces, which uses the Apertus model via the Swiss AI platform (but it's compatible with other models and platforms).

We generated both a knowledge base and tooling to:

  • Get curated information about Swiss hospitals, associations and care centres (using web scraping)
  • Get real-time information from Wikipedia articles on specific diseases (using MCP)

Then, we fine tuned system prompts and RAG templates to Apertus, to make sure it gave auditable and correct information in a format that would

Challenges we ran into

The Apertus model is in its infancy, and brought us some challenges:

  • In some cases, it refused to follow the prompt properly, or focused only on specific parts of the prompt, interpreting other parts randomly. After multiple iterations, we got to a comfortable solution. Using other models, like GPT-5 or Mistral-medium made creating a prompt much easier.
  • Inconsistency between responses: for the same prompt, we got long and short responses. Reducing the temperature helped.
  • API rate limits: sending the prompt + RAG template + input + output, and even for article summarisation made us reach the limits very quickly.
  • The model started to loop in a few cases, where it would not stop generating very similar sentences one after the other.

On top of this, finding the data and cleaning it for the model to consume was not as easy as we thought it would be, and there is still work to be done on this side.

Accomplishments that we're proud of

  • We were able to find the data and create clean-up code that would improve the results of the model
  • We found a good structure for responses that would be relevant and beneficial for the patients and families.

What we learned

Most of us were not familiar with these tools and protocols. We learned to use Open API, RAG, MCP, data cleanup, Open WebUI and prompting.

What's next for Lighthouse

We should add missing data, clean-up and improve data quality, and improve the prompts and models behind (potentially with fine-tuning), so that the answers are more correct and it adds traceability with sources.

Built With

Share this project:

Updates