I spent the last 8 months building YouClip, an AI video clipper designed to turn long podcasts into viral shorts automatically.
Most competitors burn cash on expensive GPU instances for video rendering. I wanted to prove you could build a scalable engine on dirt-cheap CPU instances.
The Result
I managed to get the entire orchestration (FFmpeg + AI Analysis) running for under $15/month in total infrastructure costs.
- Compute: $5/mo (Standard CPU VPS). No GPUs.
- AI: ~$5/mo (Gemini Flash/OpenRouter).
- Proxies: ~$3.50/mo.
The Architecture (The Secret Sauce)
I decoupled the monolith into 7 distinct microservices orchestrated by Docker & N8N.
- Frontend: Next.js 16 (App Router) for the UI.
- Brain: N8N handles the state, error handling, and directing traffic between services.
- Worker: Python FastAPI specifically optimized for CPU-based FFmpeg processing (auto-reframe, face tracking).
- Intelligence: Gemini Flash for scoring the "virality" of clips.
The Challenge: FFmpeg on CPU
The hardest part was preventing the API gateway from timing out while FFmpeg was crunching 1GB video files. I solved this by using N8N webhooks to handle asynchronous callbacks. The Next.js frontend polls the status, while the Python worker churns through the video in the background.
Open Sourcing the Architecture
I recently landed a new full-time role as an AI Engineer, so I won't have time to market this SaaS actively.
Instead of letting the code rot on my hard drive, I decided to package the entire engine and source code for other devs who want to skip the 8 months of R&D I went through.
You can grab the full architecture and starter kit here:
👉 View the Repository on GitHub
I also documented how to handle the N8N <-> Python handoff in the README.
Happy coding!
Top comments (1)
Love the transparency here. In a world of 'unlimited VC funding' posts, seeing a scalable engine built on a $5 CPU budget is incredibly refreshing. Congrats on the new AI Engineer role, too!