Aleksey Yudin

Self-directed transition into ML/NLP with public work along the way. 8 hackathons in 2025–2026, team lead in all 5 team events (2–4 people; design doc and architecture before the start). Full list with metrics and repos: gist.github.com/AlekseyYudin-161/2cf4408e53dcbeb63618242a3e269015

• JunMate — multi-agent AI résumé assistant (solo, 30-day MVP): 6 LLM agents with pydantic-structured handoffs, LLM-as-a-judge critic for factuality, eval pipeline (track-classification accuracy 90–100%), unit tests + CI, tiered model routing with cost tracking; deployed on Streamlit Cloud. Finalist of "Kodik's Choice" & top-10 "Community Choice", Kodik Launchpad. github.com/AlekseyYudin-161/JunMate

• Wink AI Challenge 2025 — 10/105, Team Lead / NLP. NER on film scripts (18 entity types, 37 BIO tags), fine-tuned bert-base-NER-Russian — weighted F1 0.917; ruT5 scene summarization.

• Dev-to-Dev: Agentic Engineering Challenge — 4/50, solo. MCP server for codebase auditing in Cursor / Claude Desktop: Python AST + regex tools (tech-debt markers, secrets, code smells), Docker, smoke tests; write-up on Habr (RU).

• Yandex ML Challenge 2026 (Effective Inference). LoRA SFT of Qwen3-1.7B on 10k pairs (completion-only loss, adapter merge), offline vLLM inference in Docker without network, local LLM-as-a-judge for hypothesis selection; score 46.0 → 56.2 (private).

Also: PurpleHack 2026 by Sber — 27/106 (LLM hallucination detection from GigaChat3-10B internal states, 257 features, LightGBM); Magnit Tech × HSE RetailHack — 6/28 (drift-aware revenue forecasting: adversarial validation, CatBoost + Optuna); MyDigest — Telegram news-digest bot, Simulator ML team project (async scraping → MistralAI API with backoff/jitter retries → Supabase). Details in the gist above.

• Built and shipped production RAG support bots in Telegram ( ≈ 1,500 users): answers to text queries and app screenshots over a knowledge base of 400+ PDFs.

• Retrieval: hybrid search — pgvector (multilingual-e5) + PostgreSQL tsvector/BM25 fused via RRF; cross-encoder reranking of the top-30 (bge-reranker-v2-m3); top-1 rerank-score threshold to refuse irrelevant queries.

• Offline eval on a 31-case golden set: Hit Rate@5 = 1.0, MRR = 0.907, Refusal Accuracy = 1.0; quality logging (latency, refusals, docs_returned) for monitoring and iteration.

• Russian text preprocessing (pymorphy3: lemmatization, synonyms, case forms), prompt-injection guardrails (15+ RU/EN regex rules); Docker Compose (bot + PostgreSQL, persistent volumes).

• Result: support workload dropped ~85% (client's estimate + logs); median response ~5 s for text queries. Sole ML engineer — full cycle from requirements to deployment and monitoring.

• Contract role; reference available on request.