The Roast

A daily digest, brewed just how you like it.

Comment

Inspiration

Every morning, people get dozens of emails from all different kinds of content sources: NY Times, Tech Crunch, Quora, etc.

What it does

Our platform let's you specify your interests and the places you get your news, then sends you a daily newsletter every morning, with news and online content curated for you.

How we built it

We used React for the frontend, and Flask/Postgresql for the backend. We used Claude and GPT-3.5-turbo-0613 and utilized LangChain to structure our models. We use Claude to rerank the news sources, and then choose the ones which will be used to write the newsletter. Since we ask our LLM's to return JSON formats, we take advantage of GPT's newly upgraded capacities to fix broken JSON formats generated by Claude.

Challenges we ran into

We found that storing the newsletters in our database was not optimal, and if we had more time, we would definitely have taken advantage of embedding database frameworks like Pinecone or LlamaIndex. Creating the code connecting the backend to the frontend, and troubleshooting API calls to see what went wrong when content didn't render on the frontend. Prompting GPT to create coherent and relevant newsletter was also a difficult task. Also, bandwidth problems were a huge issue, so much so that we ended up camping out in the currently student-occupied Anthropology Library just for decent bandwidt.

Accomplishments that we're proud of

We're proud of creating a robust interface that fully showcases the potential of an interactive newsletter, with content tailored for your interests. We are also proud of how quickly we were able to make a fully-connected frontend and backend with limited database and UI design time.

What we learned

We learned how difficult it can be to create intelligently designed databases that don't result in hour long database restructure and debugging.

What's next for The Roast

We hope expand our interface's capabilities by creating an interactive chatbot, which can point out specific parts of it's newsletter, or quote important sections of the news it has cited, to further the interactivity of the newsletter. Given our database restraints, we see using embedding databases as our next step of development.

Built With

Share this project:

Updates