Looking back, dated to the end of the prototype week.
Ratard: from a one-minute pitch to a browser extension in a week
We won Best Pitch at HPI Engine's E-Challenge with an idea against fake news. A week later I started building it.
One minute on stage, one slide, and then a sign that says “Best Pitch 2024”. That was December 13. On the 21st I opened an editor, because a pitch alone felt a bit thin.
The setting was the E-Challenge, a pitch competition inside the economics course run by HPI Engine. Around 15 groups, all on the same topic: fake news. Our idea was Ratard, a browser extension that reads along with you. You stay on the page you were reading anyway, and the extension marks what looks off.
How it is supposed to work, in plain words. You open an article. The extension goes through the text and colours passages by four scores: how likely something is fake news, how manipulative the wording is, how biased, and how far you can trust the source. Click a marked passage and an overlay opens with a fact check, a rating of the source and a warning if a link leads somewhere dubious. If you think something is wrong, you can send the passage to the backend from the context menu.
What I actually built in those six days, alone, in four commits: the WebExtension front end, about 1,600 lines of JavaScript with the highlighting, the overlay and the popup, plus a small FastAPI backend in Python. The backend is a skeleton. The scores you see right now come from placeholder logic, not from a model. The plan was Transformer models from Hugging Face for the text classification, and later images and videos.
That is also the honest status. Concept and first implementation, yes. A product that tells you reliably what is fake, no.
Two things I noticed while building. First, the UI part goes fast. Colouring text on someone else’s page, positioning an overlay, a popup, that was done in days. The hard part is the thing behind it: a model that decides what counts as manipulation and can explain why. That is where a one-minute pitch is easy and the code is not.
Second, building right after the pitch helped. All the questions we had waved away on stage came back on day one. Which sites do you even read? What happens when the backend is down? Who decides what “bias” means? None of that is on a pitch slide.
Tools and skills: JavaScript and the WebExtensions API, HTML and CSS for the overlay, Python with FastAPI, a first look at NLP classification with Transformers, and a lot of thinking about how to show a warning without being annoying.
The repository is private for now. The idea stays on my list.