Daniel Ivanov — csylabs

Research Lab · Исследовательская лаборатория

Benchmarks, weights, and methodology — in the open

Benchmarks, methodology, model weights — published as they ship. Commercial work runs on the same stack anyone can audit.

Benchmarks that didn't exist

EduBench-RU + ЛИИ-Спорт-Bench-RU (655 Q × 35 sports) — open RU domain benchmarks, live at bench.csylabs.com. Methodology open, results reproducible.

View the benchmarks

Open-weight models

Domain models on a Gemma 4 31B base — fine-tuned for education, Chuvash, and sport. Open-weight releases ship on HuggingFace as they clear our 20-item release gate; first sport model lands June 2026.

Models on HuggingFace

Published methodology

DoRA + PiSSA + rsLoRA + LoRA+ fine-tune pipeline · 4-phase PII Memorization Audit · 6-layer safety stack · Apache 2.0 / MIT licensing · 20-item release gate.

Habr engineering blog

Benchmark snapshot · May 2026

ЛИИ-Спорт-Bench-RU leaderboard

Score scale: / 10

1

claude-opus-4.7

9.10 / 10

2

gemini-3.1-pro-preview

8.88 / 10

3

gpt-5.5

8.53 / 10

4

deepseek-v4-flash

8.03 / 10

5

qwen3.5-27b

7.52 / 10

6

gemma-4-31b-itour base

7.45 / 10

Closed frontier wins raw — but isn't fine-tunable or RU-sovereign-hostable. Gemma 4 31B is the best base we can fine-tune and deploy under 152-FZ. That's what we ship.

Full leaderboard →

Latest

What's shipping

HabrТри попытки обогнать в бенче базовую Gemma 4 дообучением — и все три мимо

27 мая 2026 г.

TalkЯндекс × Дом народов III — ИТ и языки народов России (спикер панели)

21 мая 2026 г.

HabrПрогнал семь LLM через свой русский спортивный бенчмарк. Базовой моделью всё равно оставляю Gemma 4 31B

18 мая 2026 г.

HabrКак за 30 000р дообучить модель, которая работает на уровне GPT-5.4 — на задачах российских учителей

22 апр. 2026 г.

HabrМы протестировали 22 нейросети на задачах для российских учителей. Ни одна не знает чувашский

25 мар. 2026 г.

Integrator · AI integration

Three ways to engage

From a 2-week readiness audit to enterprise deployment. The same research-grade pipeline that backs our open releases — applied to your domain data.

LLM Readiness Assessment

400K ₽

2 weeks

Domain benchmark run on your data
Fine-tune scope + deployment estimate
Measured baseline against SOTA

Book an assessment

Most common

Bespoke Fine-tune + Deploy

2–5M ₽

12–16 weeks

Fine-tuned model on your domain
On-prem or Selectel 152-FZ deployment
Eval harness · 3-month warranty

Discuss a project

Enterprise Platform

80K–2M ₽/mo

Month-to-month

Your domain model served on-prem
Ongoing data curation · quarterly model revisions
Benchmark tracking · multi-tenant or dedicated

Enterprise enquiry

Platform · Платформенная инфраструктура

Eight years of video delivery and SaaS infrastructure

Live infrastructure the products on this page run on. The technical foundation our AI work builds on.

Video Delivery

VDN

Private video delivery infrastructure. Encrypted pipelines and edge distribution for controlled-access live streaming — purpose-built for distributed live events.

Sports Analytics

Insight

Federation intelligence. Player analytics, scouting, and league data pipelines — 22 leagues, 22.7M+ events, 8 seasons.

Operations

Ops

Sports operations. Event management, task tracking, contacts, and budgets for clubs that outgrew spreadsheets.

Education

Learn

Interactive language lessons with an AI teacher. Watercolor-illustrated daily sessions for Chuvash preservation — the first digital K-11 curriculum for an endangered language.

Stack

TypeScriptNext.jsReactPythonPostgreSQLHasura GraphQLbetter-authTailwind CSSGemma 4Claude APILangGraphvLLMpgvectorRAGDoRA / PiSSA / rsLoRAMistServerSRTWebRTCHLSFFmpegDockerKubernetesAnsibleProxmoxSelectelCloudflareCaddyLinuxGitHub Actions

About · Daniel Ivanov

Founder-operator, from Cheboksary

Eight years leading remote video production for international sports — twenty thousand live events across multiple countries. In April 2026 I went full-time at ООО ЛИИ — a company I founded in Cheboksary to turn that operational experience into AI products with a research foundation.

Every platform on this page is live and serving real users — built by me with a small team where AI is a force multiplier, not a replacement.

I write about the process:

Benchmarks HuggingFace csylabs Habr daniel.csylabs.com @techaroundsports

Contact · Start a project

Tell us about the domain

Drop your context — we reply within 24 hours and propose a 15-minute scoping call. For urgent conversations, email hello@csylabs.com directly.

Have a domain worth training for?

hello@csylabs.com

Open benchmarks. Domain models. Deployed under 152-FZ.