Launched an LLM integration service — starting with education
Today I published a product page at csylabs.com/llm-integrator. AI integration for Russian organizations, built on large language models (LLMs) — 60-90 days from Discovery Week to a working pilot. Starting with education. Then healthcare, then law. Here's why these industries, in that order.
What it is
Not consulting. Four stages with fixed prices and a gate decision at each:
→ Discovery Week (5 days, 200-500K ₽) — we map your processes, find where AI delivers measurable efficiency, write the pilot scope document. → Pilot (60-90 days, 2-5M ₽) — deployment on certified Selectel infrastructure, integration with your IT environment, team training, case study with numbers. → Scale Retainer (500K-1.5M ₽/month) — if the pilot worked, we scale and maintain. SLA, quarterly roadmap, onboarding for new team members. → Knowledge Transfer Bootcamp (100-300K ₽/cohort) — two-day workshop and four weeks of follow-up for faculty councils, ministries, regional operators.
You can stop at any stage. Discovery Week is often enough to understand whether a pilot is even needed.
Why this service, why now
Most AI projects in organizations die on two barriers.
First — 152-ФЗ-compliant hosting with FSTEC and GIS K1 certificates. If you build it yourself, that's 12-18 months of internal development. Most schools, universities, and hospitals can't wait that long.
Second — no methodology for measuring results. "We deployed AI — did it get better?" without measurable KPIs turns into a subjective discussion at the next meeting. A year later no one remembers the initiative.
For the past two years, I've walked through these same two barriers from the other side. Built EduBench-RU — the first open benchmark for Russian-language pedagogy, 30 LLM models, methodology with three independent judges. The first Habr article drew 5,400 reads and 30.5% read-through. Then I fine-tuned EduLLM-RU 27B for $330 — it placed #9 out of 30 on Russian teacher tasks and runs locally, never leaving Russian infrastructure.
Deploying it for schools on Selectel servers (ООО ЛИИ partner program, 152-ФЗ level 1, FSTEC orders 17/21, GIS K1 — all included in the standard offering, no markup).
Beyond 152-ФЗ — a few things clients get as part of the package:
→ Open methodology — the benchmark can be verified internally and defended to auditors → A model fine-tuned on your data — not a one-size-fits-all fixed vendor API → Independence from sanctions risk and vendor service shutdowns → Publication potential for academic partners — papers, reproducibility, co-authorship
The page is a formalization of what I'm already doing with pilot sites. Four stages, fixed prices, open methodology — so the next conversation doesn't start with "show me your reference list."
Why education, then healthcare, then law
The three most regulated industries in Russia. And it's precisely where the stakes are highest that the methodology for measuring results matters most.
Education — first. I've been here two years: open methodology (EduBench-RU), my own fine-tuned LLM (EduLLM-RU), pilot schools and universities in progress, academic co-authors in discussion, Habr articles read by both teachers and developers — each pulling their own layer. Academic infrastructure recognizes its own language — peer review, reproducibility, open data. The conversation is between peers.
Healthcare — next. The regulatory shape is very close to education: medical data falls under 152-ФЗ level 1 or 2, and the methodological work transfers directly — you can build MedBench-RU on the same three-judge principle. The barrier is higher (physician liability, medical device certification), but the stack and process are ready.
Law — third. The most text-heavy industry, where LLMs actually work: contract generation, case position drafting, practice analysis. The Russian legal tradition is formal documentation — an ideal field for automation. Clients are conservative, but once they engage, they stay.
The general logic is simple. Go into industries where everyone gets burned on 152-ФЗ and puts off AI adoption. Not into places where AI already enters easily — advertising, e-commerce, consumer services — where competition is fierce and holding position is hard. Into the complex ones, where the regulatory barrier scares others off. Which means: whoever clears it first has a longer-lived advantage.
Education is the slowest of the three by decision cycle, but it's where I can walk in with open cards: the methodology is published, the model works, academic conversations are live, teachers have been reading my material for two years.
What's next
The page lives at csylabs.com/llm-integrator. If you work in education and are thinking about AI for a school, university, or regional operator — get in touch. Discovery Week is five days and a fixed sum; it doesn't commit you to anything further. Most often the output is a 10-15 page document with a process map, an estimate of where AI delivers efficiency, and a recommendation on whether a pilot makes sense.
I write in more detail about methodology — on Habr, about the daily work — in @techaroundsports.