Chapter 4 of 8

Infrastructure we control

What JT Digital built since March · 3 min read

The heavy work behind JT Digital's apps runs on machines we control: video renders, face detection, language models, a web crawler and a browser that reads web pages. A message queue hands out the jobs, and the files the machines read and write sit in storage in the same datacentre in Germany. Every app reaches this work through the queue.

Every job goes through one queue to machines we control, next to the storage that holds their files.

The Everyframe engine is the biggest user. It renders every video for jt-cut, finds the faces huluma covers and renders every ad version Everyframe Composer makes. Planet Positive Sport sends each committee's objectives to a language model here, which links them to the UN Sustainable Development Goals. Kataster runs its crawler and its company profiles on the same machines, and StorySell uses the browser service to read the web pages it needs.

Language models we host ourselves

The machines with graphics cards run five open-weight language models, which means anyone may download them and run them on their own hardware. Each one has a licence that allows commercial use, and nothing is billed per token, the unit these models count text in. When a person is waiting for an answer, their job goes ahead of background work, so nobody waits behind a batch of company profiles. The same machines run a model that adds colour to black-and-white footage.

Long jobs that finish exactly once

Long renders were the first thing to break. The queue treats a worker that holds a job for 30 minutes as dead and hands the job to another one, so our longest renders started over and ran twice. We moved the ownership of every job out of the queue and into a record we keep ourselves, and now a job that runs for an hour still finishes exactly once. We wrote up how the job runner works on our how-we-build pages.

A browser for our apps

Since 13 September the machines also run a browser service. An app asks it for one web page and gets the page back. Before the browser connects, it checks the address, and it refuses any address that points inside our own network.

A crawler for company websites

Kataster's crawler has run on these machines since 4 September. It opens a company's home page in a real browser and fetches the site's other pages directly, one request at a time and two seconds apart. The pages it collects go to the file storage beside the machines.

Next · Chapter 5 of 8

Everyframe Composer, our main product

1 min read →

4 / 8 Infrastructure we control Chapters

Stay in the loop

Don't miss the next one.

One email when we publish something worth your time.

No spam. Unsubscribe in one click.