Your AI sidekick for content that clicks. Get AI-powered title suggestions and automated topic chapters for your YouTube videos hassle-free.
- YouTube videos and content in general has to capture a user's attention in mere seconds. A title is the first thing (or second after the thumbnail) that the user reads.
- YouTube videos with topic chapters improve retention. When users see chapters, they can quickly jump to the sections that interest them, keeping them engaged.
- I wanted to fine-tune an LLM, and mess around with it.
- Title Generation
- Current AI models can generate titles for you, but it's obvious they are AI-generated. To fix that, I fine-tuned Gemini Flash using this dataset https://github.com/chris-lovejoy/youtube-titles-and-transcripts, making it smarter and more natural at crafting titles that click.
- Upon fine-tuning, here's the result I got.
- The titles generated by this fine-tuned AI model were way better than the native Gemini model.
- The generated titles got even better after prompt engineering.
- Chapter Generation
- Just like with title generation, I wanted a model that’s better at topic segmentation and chapter creation.
- To achieve this, I fine-tuned another model using YouTube transcript data.
- I scraped YouTube using the YouTube Data v3 API, feeding the model text transcripts as inputs and segmented paragraphs as outputs.
- Unfortunately, the model didn't perform well and converged at a loss of 10.0.
- The results were clear, Gemini is better at topic segmentation natively, so with a bit of prompt engineering, I started seeing good results
- Database: GCS for storing models, training datasets, and validation datasets
- APIs: Youtube Data v2, SupaData (For extracting transcripts)
- Backend: Next.js API Routes, GCP Vertex AI (Where I deployed the two models, one fine-tuned and one native)
- Frontend: Next.js, React, Tailwind CSS, ShadCN UI
- youtube-transcript (npm) worked locally but failed in production due to Google blocking cloud IPs.
- Tried a Google Cloud Run function with a Python package—blocked.
- Set up a proxy server on DigitalOcean—also blocked.
- Solution: Used Supadata, which allows 100 free transcript requests per month.
- AWS Sagemaker? Realized it’s not free.
- Hugging Face + T4 GPU (super cheap per hour cost) worked for training, but deployment was difficult.
- Solution: Used Google Cloud Platform for training and deployment with free credits.
- Enhancing the accuracy of timestamps.
- Allowing users to upload audio/video (another challenge cost-wise).
- Deploying the third feature, that lets users upload their video and generate shorts from it.