LifeLine

Inspiration

Services are hard to find because of hard text and many sources. We believe that hard text and tricky websites should not be a barrier. We use EasyRead (https://www.mencap.org.uk/easy-read) text in our app to improve accessibility. We finetuned a Large Language Model (LLM) to produce EasyRead-style text. We collected lots of services. We made everything very easy to find and understand.

What it does

LifeLine collects services in the Edinburgh area and makes them easy to find. We also make each service easy to understand. We also made it easy to find a general practitioner (GP), get a Young Scots card, and find student deals. LifeLine saves you money and keeps you warm.

We included these services:

free menstrual products
warm places
NHS buildings
food banks
pharmacies
public toilets

How we built it

We finetuned the LLM by distilling from a larger model and correcting the output. We built the web app with Python and FastAPI for the server, and plain HTML, CSS and JavaScript for the app. We used a small model, Qwen2.5-1.5B, and taught it EasyRead with LoRA. LoRA trains only about 1% of the model, so it was quick to train. The model runs on a normal laptop, with no paid API. We collected 559 services from public data: NHS Scotland open data, Give Food, the Great British Public Toilet Map, My Period and City of Edinburgh Council. We stored them in a SQLite database. The map uses Leaflet and OpenStreetMap. Read aloud and live captions use the speech tools built into the web browser. We published our model on Hugging Face so others can use it.

Challenges we ran into

Some services had no easy way to get their data. One site had no API, so we read the data from its web page. Food banks did not list their opening hours. Our server crashed when two people used the model at the same time, so we made it do one job at a time. The model sometimes added facts that were not in the text, or changed the meaning of legal words. So we always show the original text too.