Four months. One game. Real code running for a real client.
Every summer we run an internship at Brigada. This year it started with a challenge: we built an AI bot, gave it one secret word, told it never to say it, and put it online. 38 people signed up. 26 tried. 21 got the word out, and the shortest trick that worked was eight words long.
Then we flipped it: build the bot, and let everyone else try to break yours. 16 people did, and each bot took 23 attempts. The full rules are in brigada.dev/journal/ouroboros.
The top four got a seat with us. Eight weeks of coursework on one question, how does a computer understand what you say, from simple rules all the way to training your own model. Two of them came into the office every single day, so those two got the real project.
A client and partner of ours wanted their own AI model: one that runs on their own servers, answers faster, and never sends what their users say to anyone else. The way you build that is distillation, where a big model teaches a small one until the small one can do the job by itself.
The client recorded around 1,400 voice commands in German and English. The interns cleaned the data, trained the student, and proved it worked: same recordings, same small model, a better teacher. Accuracy up 9.5 points, mistakes nearly cut in half.
In numbers: 136 commits, 31 pull requests, 15,000 lines of code, in under a month. Endri Emini and Rigon Shahini finished the whole programme and were certified in August. What they built is running in the client’s staging environment today, and Rigon is still shipping to it.
That's one summer at Brigada. We build products, fast.