Overview
A one-shot brainrot generation pipeline that transforms single prompts into complete short-form videos. Designed to lower the barrier to entry for content creators looking to produce TikTok-style animations and videos.
Demo Videos
What It Does
- End-to-end generation: From a single prompt, the system generates scripts, directs scenes, and produces finished videos
- Programmatic video creation: Uses Remotion (React-based library) to manipulate and animate assets into cohesive videos
- Wikipedia to TikTok: Can transform Wikipedia pages into engaging short-form content
Technical Challenges
- Running open source SOTA video understanding models to analyse and label memes videos (understanding meme sentiment)
- Experimenting with different LLM models and methods to cost optimize whilst maintaining quality output
Automated AI Labelling
Built a system for automatically labelling meme videos to use as video assets, including experimentation with video understanding models.
Unit Economics
Focused on keeping cost per video low while maintaining quality:
- Implemented prompt caching
- Evaluated cheaper models (Gemini vs Claude Sonnet)
- Used code where possible instead of AI (such as RAG for asset retrieval)
Production Quality
Balancing the cost of production with the quality of output, ensuring the best possible LLMs are used where they matter most.
Tech Stack
- LLM: Claude and Gemini for main generation
- Voice: ElevenLabs for narration synthesis
- Audio Processing: OpenAI Whisper for timestamps
- Asset Retrieval: RAG with SentenceTransformers for embeddings
- Video Compilation: Remotion (React/TypeScript)