The Problem Knowledge creators often already have the most important part of a video: the idea and the script. Producing a professional biographical or investigative documentary traditionally takes over 5 days .The real bottleneck is production. Turning one script into a finished knowledge video requires research, scripting, visual selection, editing, camera movements, voiceover, subtitles, and platform formatting. Creators spend hours connecting different tools and performing repetitive tasks that add little creative value. We wanted to build an AI agent that could take a creator from idea to publish-ready video.
Our Solution We built an end-to-end agentic workflow with five phases: Research → Scripting → Visual Curation → Video Production → Distribution The workflow first uses Google Search and Gemini to research and deconstruct a topic, then generates a structured storyboard and script. An automated quality audit and approval gate ensures the content is ready before production. Once approved, the system intelligently selects visual assets, normalizes them, and uses Gemini's multimodal vision capabilities to verify their quality. A dynamic camera engine powered by FFmpeg then creates motion and seamlessly assembles the video. Finally, Neural TTS generates the voiceover, millisecond-level SRT subtitles are created, and the system produces a mastered H.264 video ready for multi-platform distribution.
Why an Agent? Instead of using separate AI tools for research, writing, editing, voiceover, and subtitles, our workflow connects these tasks into one continuous production process. More importantly, the system does not simply generate a video from scratch. It understands the script, reasons about what visual content is needed, selects and transforms existing assets, and validates outputs at critical stages. This makes AI video creation more practical for knowledge creators who want to reuse real-world footage rather than generate every visual with AI.
Challenges & Lessons Our biggest challenge was balancing automation with quality. Errors early in the workflow can propagate into the final video, so we introduced reflection and quality gates at key stages. We also learned that effective AI video creation is not about generating more content. It is about orchestrating the right decisions at the right time. Our goal is to let creators spend less time operating production tools and more time on what humans do best: developing ideas, building expertise, and telling meaningful stories.
Log in or sign up for Devpost to join the conversation.