How legit2clip started

A friend's question

A friend asked if I could take his hour-long cybersecurity talks and turn them into clips, to promote his next talk.

On the drive home I worked out the steps:

1. download the talk;
2. transcribe every word;
3. find the moments worth sharing;
4. cut and frame them;
5. burn in the captions.

One big program

The first version was one program: 55,000 lines of Python. It worked, but only for me, one video at a time. It couldn't take work from anyone else.

So I added accounts and sign-in. Then I split it into small services, so each step could grow on its own.

The server died

I went to upgrade my network to 10 gigabit, and my main server died.

I moved to the cloud, but cloud prices with no customers and no time to grow didn't add up. So I rebuilt on a stack of small mini PCs.

I kept building meanwhile: Synestia, our AI music app, and some model-training projects.

A network instead of a server farm

When my server came back, I picked the clip project up again. By then it had outgrown one machine: clips, transcripts and AI models all wanted graphics cards at the same time. First I made the scheduler work across projects.

I couldn't afford the graphics cards myself. But I could build a network where people with spare ones run the work, and get paid for it.

That network is legit2compute, and all of my own work runs on it now.

100 hours of real talks

Our first outside user came on board, and a spare machine joined the network.

We ran 100 hours of talks from SecKC, Kansas City's hacker meetup, through it: their whole archive of 156 talks, some of them several times over. We found the bottlenecks and engineered around every one. Here's what we learned.

The beta is live

Twelve months after that first question, legit2clip takes a long video and hands back the moments worth sharing, cut, framed and captioned. You see the price before anything runs.

I'm looking forward to the next twelve months, and I'd love you to be part of them.