Python / AI developer building end-to-end ML and data systems - RAG pipelines, LLM fine-tuning, NLP, and production scrapers. Not notebooks: things that run, are tested, and ship.

veriscope - news fact-checking assistant (v1.0 released) Decomposes a news article into atomic, checkable claims, retrieves web evidence for each, and reports honest per-claim verdicts with citations - deliberately no fake "truth score".

• Full pipeline: claim decomposition → per-claim evidence retrieval → stance detection (supports / refutes / NEI) → verdict aggregation with manipulation-signal flags; source-independence check: 20 reprints of one press release count as one source
• Measured, not claimed: 83% verdict accuracy on a 75-claim calibration set (no fabricated story was ever confirmed), plus a stratified AVeriTeC benchmark with documented failure modes
• FastAPI backend + Manifest V3 browser extension · Qwen2.5 · pgvector · Docker · CI/CD with tagged releases · live demo

gazeta-lora-eval - LoRA fine-tuning + automated eval harness QLoRA fine-tuning of Qwen2.5-1.5B for Russian news summarization, with a harness that objectively compares baseline vs fine-tuned on a held-out set — not training loss, but ROUGE-L, BERTScore, and pairwise LLM-as-judge.

• Found and fixed 3 measurement bugs the naive setup hid: prompt truncation cutting the response marker, a verbosity-biased judge, and a repetition penalty that drove the baseline into English
• Honest result: quality parity with a strong instruct baseline at −34% summary length and better automatic metrics
• Single shared prompt template across train/eval/serve · completion-only loss · FastAPI serving both variants from one weight copy via disable_adapter()
• Stack: PEFT · TRL · bitsandbytes · transformers · FastAPI

RAG-chatbot-API - self-hosted RAG with citations Fully local RAG: upload PDF / DOCX / Excel or point it at a URL, ask questions, get answers grounded in the source with citations.

• Hybrid retrieval (vector + full-text merged via RRF) with local cross-encoder reranking
• Follow-up question rewriting, OCR fallback for scanned PDFs, streaming responses (SSE)
• FastAPI · LangChain · Postgres/pgvector · Ollama (no API keys, runs on CPU) · Vue 3 · Dockerized · unit-tested · CI · automated answer-quality eval

vatoring - full-stack Telegram Mini App A GitHub-style Mini App to monitor your GitHub/GitLab repositories, CI status, and assigned issues — authenticated with Telegram.

• React + Vite + TypeScript frontend, FastAPI backend, aiogram bot, Docker Compose (Postgres / Redis)
• Server-side Telegram initData validation, per-user provider tokens · deployable to Railway

market_parser - production scraping pipeline Collects daily prices across 19 Russian retailers/marketplaces with anti-bot evasion.

• Camoufox / Playwright with a hybrid cloud+local architecture for the hardest, captcha-protected targets
• Scheduled runs, export to SQLite / XLSX / Google Sheets, Telegram delivery · deployed on Railway with cron

Applying to AI / Computer Science bachelor's programs abroad while going deeper into ML/DL fundamentals (model internals, PyTorch). I learn best by building real, running systems — and I want my GitHub to reflect that rather than tutorial-following.

Python · PyTorch · PEFT / LoRA · FastAPI · LangChain · PostgreSQL / pgvector · Docker · Playwright · React / TypeScript · Git / CI

Telegram @vakudo