Platform
Seedflows Training
Fine-tuning centre for custom image models (LoRA) on serverless GPUs. The product backend never talks to a GPU vendor directly.
The problem
Training custom image models needs expensive GPUs for a few hours at a time, and that should not leak into the product that uses the models.
What I built
- A separate training service that starts LoRA runs on Modal and only pays for GPU time while a run is going.
- A proxy in front of Modal, so the main product never talks to a GPU vendor directly.
- A management backend and UI to start and follow training runs.
Stack
- Python
- Modal
- Spring Boot
- Angular
- Docker