Adding some detail on how we used a few of the MLH sponsor tools.
ElevenLabs powers our voice A/B testing. A creator gives us a base video and a script, and we generate several voiceover reads with ElevenLabs using different premade voices and voice settings like speed and stability. We mux each read onto the video with ffmpeg and then score every version. This is the one piece of content we actually generate, and it lets a creator A/B test how new voiceover affects engagement without recording anything themselves.
MongoDB Atlas is our system of record. Every user, test, variant, and score lives in Mongo, accessed only from our Python backend through pymongo. Our history view and our personalization both read from these collections, so Mongo is what lets the product remember a creator's past tests instead of treating each one as a fresh start.
Gemini runs our suggestion engine. We send the base video to Gemini as a multimodal input so it actually watches the footage, the pacing, the first few seconds, and whatever audio is already there. Based on what it sees, it proposes voiceover variants worth testing, with specific scripts, speeds, and tones that fit the video. We picked Gemini specifically because it can see the video, so the suggestions come from the real content and not generic advice. We route it through OpenRouter after our direct key kept rate limiting during the hackathon.
Auth0 handles login. The frontend uses the Auth0 SPA SDK, and our backend verifies the access token against Auth0's JWKS before trusting any request, which gives us a stable user ID per creator. We moved to Auth0 after finding that our frontend platform's built-in login could not be verified by an external backend on the free tier, which would have quietly broken the link between our frontend and our backend.
Log in or sign up for Devpost to join the conversation.