Diploma in software development — Ust-Kamenogorsk Higher Polytechnic College, 2026; currently studying towards a bachelor's degree.
I work on what sits underneath the model: annotation, its measurable quality, and the acceptance of work done by others.

Kazakhstan  ·  open to relocation  ·  email  ·  telegram

Five annotation types, each annotated by hand in CVAT blind to a public ground truth, then compared against it by the metric that type is actually measured with. Every repository carries its own annotator guidelines, a report on the systematic disagreements, and a reproduce block.

_{The recurring finding is that most of the disagreement is convention rather than carelessness: the frame border costs 72 of 77 uncovered boxes on tracks, letter case costs 94% of the character error rate on text, and the visibility flag is an axis no coordinate metric can see at all.}

Acceptance of someone else's annotation is a sixth, private stage: a sampling plan with a stated risk, three defect levels, an accept-or-return rule, run against two batches — a human-made reference and a model pre-annotation.

steel-defect-classification — six defect classes of hot-rolled steel strip (NEU Surface Defect Database). ResNet-18 fine-tune, YAML-configured runs, tagged release with weights. _{Validation accuracy hit 100%, so I stopped and wrote down the five pipeline bugs behind it instead of publishing the number. Honest metrics land once the split is rebuilt.}

dokuwiki-rag — retrieval over the internal DokuWiki of a metallurgical plant, built as my college diploma project. Hybrid search in Qdrant with bge-m3 embeddings and a bge-reranker-v2-m3 cross-encoder, qwen3:8b through Ollama, FastAPI + Streamlit, four services under Docker Compose. Retrieval Hit@10 = 97% over 34 hand-labelled questions; the customer's corpus is replaced with a synthetic one.