nbasearch

NBASearch is a video search engine that analyzes game footage to generate highlights, summaries and enable in-video search.

Comment

Inspiration

I already played with search based on text for other projects that I made, so I talk with Daniel from the Warriors and he told me that they have a lot of videos of games but is always complicated to search on those videos.

What it does

nbasearch is a video search engine for game footage. It analyzes videos to automatically generate text summaries, create highlight clips, and enable in-video search using generated transcripts.

How we built it

Is made with next.js, tailwind and mongodb atlas. Audios are hosted on cloudinary. All the transcripts were made using whisper on replicate.com And the summaries and highlights using respell.ai with claude instant. Search per word is working thanks to mongodb atlas indexing.

Challenges we ran into

Is hard to do background processes because vercel has a short timeout on APIs, Also the official whisper api does not have segments, it was a problem also on respell so I ended up fixing it with replicate.com

Accomplishments that we're proud of

Finish all the basic stuff that I wanted in one day basically and figuring out with a bunch of services.

What we learned

Is hard to do background process on vercel, is cool to focus and code stuff that you are already familiar with to take advantage of the time

What's next for nbasearch

I would love to do the search globally and also not only use text for search but try to understand the images of the games. I think just to get better results I just need to analyze other data, not only the audio.

Built With

Share this project:

Updates