TLDR.AI

Don't sign your life away.

Comment

Inspiration

Everyone is familiar with the hundreds of pages full of legal jargon that pop up every time your try to sign up for a website. It's long, complicated, and lets be real here, no one is reading it. But, just because you aren't reading it, doesn't mean its not important. In the growing age of data privacy concerns, obscure arbitration clauses, and terms that are WAYYY too unfair towards the user, you DESERVE to know what you're signing!

That's where TLDR.AI comes in!

What it does

TLDR.AI is a google chrome extension that reads terms of services on account creation pages, locates risky clauses and potential faults, and produces an easily digestible summary of the dangers of signing!

The extension actively runs in the background of the browser, scanning webpages for links to Terms of Services that the user is expected to sign. Once TLDR locates a link, it utilizes natural language processing AI to produce a list of dangers, a brief and simple description, and a direct quote from the TOS.

How we built it

Backend The backend consists of processing the TOS, and producing a list of accurate faults. We did this by using a wide variety of generative AI; first, we used Cohere and GPT4o to create an artificial dataset of potential faults in TOS documents, what we would want the AI to respond with, and an example quote from a sample TOS which would trigger detection. This dataset was then used to Finetune Cohere's Command R+ model to increase accuracy of identifying and reducing hallucinations. We also did significant prompt engineering to produce a system prompt which creates responses grounded in the exact TOS text. We then created a REST API using FastAPI, allowing for the TOS text to be posted into the backend, processed, and the returned to the frontend. Given the TOS text, our Cohere model uses Retrieval augmented generation to produce results grounded in the text, and utilizes Cohere JSON structured output formatting to produce a JSON object which is constrained to the exact fields we require.

Frontend The frontend is built fully on raw HTML, CSS and JavaScript. We started off with creating our initial Figma design, applying for the Figma Dev student package, and then using their VS Code integration to build our design! On the client-side the extension is constantly scanning website's URL's and body text for hints of a TOS page, and once it is found, we scrape the document and send it to our REST API for processing! Once we receive the response, we display it in an easily accessible sidebar!

TLDR, the power of friendship and redbull.

Challenges we ran into

"Separating TOS from HTML Bullshit." - Pooria Ahmadi, January 25th 9:36PM, as I'm writing this post.

INTEGRATION IS HARDDD. Even once we got our individual parts functioning it took a significant amount of time to connect all the pieces, not to mention, the large corporations who write these obnoxious TOS make it intentionally hard to easily access it, so, we went out of our way to come up with creative webscraping techniques to overcome these challenges.

Accomplishments that we're proud of

Top accomplishments:

  • Coding an animation in HTML (don't ask).
  • Our beautiful chrome extension design fully functioning
  • Leveraging EVERY feature Cohere offers to create the best results possible

Creating a product that every member of our team and every person we have talked to believes could actually be a useful and impactful product outside of the hackathon!

What we learned

We learned a lot about LLMs, Webscraping, APIs, and especially a lot about designing and creating frontends with Figma.

What's next for TLDR.AI

Publishing on the Chrome store! We hope that everyone will feel one bit safer knowing they're always getting the TLDR of what they're signing.

Built With

Share this project:

Updates

Submission history