Eskadra Bielika: teaching people to wire a Polish LLM into real systems

TODO: one or two sentences on how I ended up on the trainer side of this. How I got into the Bielik community, why I did the trainer certification in 2025, and what a test engineer who spends his days on automotive hardware and Go services wanted from teaching people about a Polish language model.

PHOTO TODO /images/blog/eskadra-bielika/01-venue.jpg TODO: venue / opening session
TODO: where and when it was

What Eskadra Bielika actually is

Bielik is an open Polish language model built by the SpeakLeash community. Eskadra Bielika (“Bielik Squadron”) is the practical, hands-on side of that community: a series of workshops where people sit down with the model and wire it into something real instead of talking about it.

TODO: the edition(s) I trained at, what tracks were offered (e.g. RAG, agents, MCP, fine-tuning, cloud deployment), how the day was structured, and what the trainer role actually involves.

What I worked on

TODO: pick the one thing you actually built. Suggested structure:

  • The problem. TODO: what was the input data, who was the imagined user.
  • The stack. TODO: which Bielik variant, how it was served (vLLM? Ollama? hosted?), which retrieval/agent framework, where it ran.
  • What broke. TODO: the honest part. Tokenisation quirks, context length, retrieval quality on Polish text, tooling that assumed English.
  • What worked better than expected. TODO.
PHOTO TODO /images/blog/eskadra-bielika/02-workshop.jpg TODO: laptops, whiteboard, people mid-workshop
TODO: caption
TODO: a small snippet worth keeping, e.g. the prompt template that finally
behaved, the MCP tool definition, or the docker-compose that served the model.

The embedded/backend engineer’s view

This is the part I care about most, because it’s where Bielik stops being “AI news” and becomes plumbing:

  • Sovereignty is a feature, not a slogan. A model trained on Polish data, with open weights, that I can run on a box in a rack in Poland. That matters for the same reasons I self-host everything else.
  • The interesting work is around the model, not in it. Retrieval, evaluation, guardrails, latency budgets, the boring pipeline that feeds it fresh data. That’s backend engineering with a new kind of dependency.
  • Small models and small hardware get along. TODO: if I tried a quantised variant on modest hardware, note what ran where and how fast. If not, cut this bullet.

TODO: one honest paragraph on where I think this fits into KhazLabz work. The point of the “AI infrastructure and pipelines” service is exactly this: keep the model close to the data, on hardware you control, and treat it as one more component with an SLO.

PHOTO TODO /images/blog/eskadra-bielika/03-people.jpg TODO: group photo or a conversation
TODO: caption

People

TODO: a few names worth thanking (with their permission), the organisers, whoever fixed my CUDA drivers at 11 pm.

Takeaways

  1. TODO
  2. TODO
  3. TODO

If you were there and remember it differently, or you’re thinking about putting Bielik into a product and want a second pair of eyes on the infrastructure side, say hi.