SpandTranslate

Translating a manga chapter is manual work: every bubble is read, translated, the original lettering is erased and the art repaired, then the translation is typeset back in. A single chapter takes hours. I wanted it to take one click.

All the heavy work runs on the server: image processing on GPU servers, translation with AI models. Closing the tab does not stop the job, and only successfully translated pages are charged.

We did not use an off-the-shelf detector for the bubbles; we trained SpandZero on our own dataset.

The architecture scales horizontally: new GPU workers spin up on their own as work comes in and step back when the load drops. A single user can send 5 GB of images, 600 chapters, 18,000 pages at once; the system slices it up across workers, so nobody waits in line.

• Small jobs stuck in line. Short jobs waited behind heavy ones. I added priority lanes; under the same load three clean-up jobs went from 124 s to 42 s.
• A crashed server. When a GPU worker died its jobs hung. Now a live worker takes them over in about 45 s, down from 6 minutes.
• Money safety. The same job is never charged twice and empty pages are never charged, guarded by automated tests.

A long job should never depend on the user's browser tab. The longer a job runs, the more likely it is to stop halfway, so I built every step to resume where it left off.