Skip to content
View ArthurBabkin's full-sized avatar
😉
Enjoy and develop every day
😉
Enjoy and develop every day

Block or report ArthurBabkin

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ArthurBabkin/README.md

Arthur Babkin

I train language models and build products on them.

Research engineer in the R&D team at T-Bank (T-Tech): post-training and alignment of the T-gen family, T-Lite and T-Pro. Before that, two years taking AI products from an idea to a release: a tutor, a lawyer, an investment analyst.

GMT+3 · Innopolis — Kazan


Now

RL environments built from scratch for agentic, tool-calling, multi-turn and instruction-following setups. SFT, GRPO, RLVR, reward design with explicit reward-hacking mitigation, and multi-agent synthetic-data pipelines on verl, vLLM and SGLang.

116M — requests a month go through the in-house models I post-train
69.6 — in-house Arena, against 65.8 for a baseline 7× larger
28.2% — acceptance rate of the track that took the paper, 263 of 931

From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix — EMNLP 2026, Industry Track. We consolidate traffic from 200+ internal applications onto a single self-hosted 32B model: a separate GRPO expert per axis, merged with two-stage SLERP, because each axis hacks its reward in its own way. Read on arXiv →


Before that

Deeplit · founder and CEO, 2023–2025 — an AI maths tutor. Raised $25 000, hired a team of 7, shipped an MVP to 2 000+ users, and took 1st place at KIvO 2024: the top education innovation in Russia, out of 1 100+ entries, named by HSE University and Alfa-Bank.

Croissan Studio · NLP engineer, 2024–2025 — a financial-analysis bot on three LangGraph agents, and a legal-contract system over 413K+ fragments of Russian law: IVFFlat → HNSW, 760 rps at 98% recall.


Selected repositories

Parimate — a video habit-report verified without a moderator: ASR, face recognition, deepfake detection, fuzzy phrase matching. Audience prize at AI Workshop Week 2025.

risk_tech_hack_akbars — 1st place, Risk-Tech ML Hackathon at Ak Bars Bank, 2025. RMSE 21 473 against 29 812.

ADMET_2024_Hackathon — 2nd place: predicting how dangerous a molecule is to a human.

stud_startup_bot — the grant-application assistant I built after a talk. 300 monthly users on a 3-day MVP.


Stack

Python PyTorch Transformers TRL verl vLLM SGLang LangGraph LangChain FastAPI PostgreSQL Airflow Docker


Elsewhere

Hugging Face · LinkedIn · Telegram · Email

Pinned Loading

  1. ADMET_2024_Hackathon ADMET_2024_Hackathon Public

    Hackathon ADMET 2024 - evaluating the danger of the molecules for humans using ML.

    Jupyter Notebook 4 4

  2. Parimate Parimate Public

    A Telegram bot for validating audio and video content using CV models, SR models, and VLMs, with deepfake detection leveraging metadata analysis.

    Python 7 1

  3. project_ml_course project_ml_course Public

    The project, where LLM with few shot prompting checks solutions of mathematical problems according to certain criteria and recommend courses to improve knowledge using ML model Catboost. Synthetic …

    Jupyter Notebook

  4. stud_startup_bot stud_startup_bot Public

    Startup Bot is a Telegram bot built to support student startup projects by helping users navigate grant applications to "Студенческий Стартап" by "Фонд Содействия Инновациям" and improve their chan…

    Python 1

  5. N4RLY/distributed-image-classifier-on-kuber N4RLY/distributed-image-classifier-on-kuber Public

    Python 1

  6. Portfolio Portfolio Public

    The repository with my diplomas, certificates and etc.