Language
without borders.
Xetun Bavori studies how machines read, parse, and generate human language — and passes that knowledge on through hands-on practice rather than passive instruction.
Every workshop on this platform is built around a concrete task: classify a dataset, tune a tokenizer, evaluate a model's output. The work shapes the understanding, not the other way around.
How this platform came to exist
Xetun Bavori started as a small reading group inside a Kharkiv university research lab. The group was working through early transformer papers and found that standard courses covered concepts but skipped the implementation details that mattered most for practical use.
The gap between a published model architecture and a working data pipeline turned out to be the real obstacle for most practitioners. That specific problem is what shaped the curriculum here.
Workshops are structured around three stages: understanding what the algorithm does, running it on a real dataset with observable outputs, and adjusting parameters until the results change in a predictable way. Repetition across those three stages builds the kind of working familiarity that reading alone does not.
What the curriculum
actually covers
NLP is a broad field and most introductory materials either stay too abstract or jump straight into framework APIs without explaining the reasoning behind the design choices. The workshops here take a narrower path: each one picks a specific technique and works through it in enough depth to answer the question of why it behaves the way it does.
Topics include tokenization strategies and their effects on vocabulary size, sequence labeling with conditional random fields, attention mechanisms compared step by step, and evaluation metrics beyond accuracy — precision at different recall thresholds, span-level F1 for named entity tasks, BLEU and its known shortcomings.
The workshop format
- Each workshop opens with a defined, measurable problem — not a topic overview.
- A short reference section provides the minimum background needed to attempt the first exercise.
- Step-by-step assignments run in order, each building on the previous output.
- A reflection prompt at the end asks participants to document what broke and what they would change.
- Exercises use openly available datasets — no proprietary data required
- Code is written from scratch before any library shortcut is introduced
- Participants keep all work artifacts after the workshop closes
- Feedback is collected after each assignment, not only at the end
People behind the curriculum
The instructional team is small by design. Each workshop is authored by one person who has used the technique in a real project and can describe where it failed, not just where it worked.
Content goes through a review pass focused specifically on the exercise sequence — checking whether the steps produce observable, interpretable results at each stage and whether the instructions are clear enough that someone working alone at a different time zone can follow without needing clarification.
- Curriculum reviewers include practitioners from computational linguistics, ML engineering, and applied research
- Workshop content is updated when a referenced tool or paper receives a significant revision
- Participant questions submitted through the platform are reviewed and used to improve exercise wording
Specializes in sequence modeling and evaluation methodology. Developed seven of the current workshop modules.