With one process where the outcome is easy to measure — support triage, document search or report drafting. We pilot it on your real data before any large commitment.

Can our data stay inside our infrastructure?

Yes. We deploy open-weight models on your own servers or private cloud, so no data leaves your environment.

How do you measure quality?

With a test set built from your real cases and automated evaluations on every change, so you know whether a new prompt or model is better before it ships.

Do we need a big dataset?

Usually not. Most business integrations use existing documents and systems through retrieval, not model training.

What if the pilot doesn't work?

You get the prototype and a report with the measurements. If the agreed criteria aren't met, we won't talk you into building it.

Do we own everything you build?

Yes. Code, Terraform, Helm charts and runbooks live in your repositories from day one, and cloud accounts stay in your name. If you ever take the work in-house, the handover is documentation and a call, not a rescue project.

AI integration for products and business processes

We connect language models to your data, tools and workflows, then run them like any other production system: monitored, secured, versioned and on call. Led by our co-founder Alexander Kaloshin.

Talk to an engineer Get a free use-case review

An engineer replies, not a salesperson — within one business day.

Pilot on your data first
Private, self-hosted models
Quality measured on your cases
Led by a co-founder

Official cloud partners

Yandex Cloud
Selectel
Timeweb Cloud

among our clients

Novotek
RZD
SberCorus
Otus
Besttrip
JedAI
IMSystems
Builddocs
Headframe
RTDS

Case studies →

symptoms.log

When teams come to us

The demo never reached production

It worked in a notebook; then integrations, access rights, quality and cost got in the way.

Security says no to external APIs

Your documents and customer data can't be sent to a third-party model.

No way to measure quality

Answers are inconsistent, and nobody can tell whether a change made things better or worse.

Costs grow faster than usage

Inference bills and latency are climbing, and nobody owns them.

scope.md

What we do

Use-case discovery

Where AI saves real time or money — and where it doesn't. We tell you both.

RAG over your data

Answers grounded in your documents and databases, with sources and access rights respected.

Agents and tool use

Integrated with your internal APIs, CRM or ERP, with least privilege and a human approving irreversible actions.

Private model deployment

Open-weight models on your servers or GPU clusters, so no data leaves your environment.

Evaluation and guardrails

A test set from your real cases, run on every change, plus monitoring of quality, cost and latency.

LLM delivery pipelines

Model and prompt versioning, rollouts and rollbacks — the same discipline as the rest of your stack.

What you keep

Use-case assessment with effort and risks
Production service with API, auth and logging
Evaluation suite built from your cases
Cost and latency dashboard

Stack

Qwen
DeepSeek
Llama
vLLM
Kubernetes
GPU
Go
Python

pipeline.yml

How the work goes

After every stage you decide whether to go further.

01 · discovery1–2 weeks
Discovery
Pick one process with a measurable outcome, check the data and constraints, agree success criteria.
Outcome: Architecture, pilot estimate and a running-cost forecast
02 · pilot3–6 weeks
Pilot
One use case on your real data, with a limited group of users, measured against today's baseline.
Outcome: A working prototype and a report. You decide: build, adjust or stop
03 · build1–3 months
Production
Integrations, auth, logging, load testing and monitoring of quality and cost.
Outcome: A production service, with code and infrastructure in your repos

git log --cases

Practice lead

Alexander Kaloshin

Co-founder

20+ years in IT. In 2014 he founded Last.Backend, a container orchestrator built at almost the same time as Kubernetes. Leads our AI practice.

Last.Backend: container management and DevOps as a service
Fault-tolerant infrastructure: Kubernetes, CI/CD, bare metal
Go engineering
Speaker at ProIT Fest and GolangConf

FAQ.md

FAQ

Where do we start?: With one process where the outcome is easy to measure — support triage, document search or report drafting. We pilot it on your real data before any large commitment.
Can our data stay inside our infrastructure?: Yes. We deploy open-weight models on your own servers or private cloud, so no data leaves your environment.
How do you measure quality?: With a test set built from your real cases and automated evaluations on every change, so you know whether a new prompt or model is better before it ships.
Do we need a big dataset?: Usually not. Most business integrations use existing documents and systems through retrieval, not model training.
What if the pilot doesn't work?: You get the prototype and a report with the measurements. If the agreed criteria aren't met, we won't talk you into building it.
Do we own everything you build?: Yes. Code, Terraform, Helm charts and runbooks live in your repositories from day one, and cloud accounts stay in your name. If you ever take the work in-house, the handover is documentation and a call, not a rescue project.

Let's talk about your infrastructure

Tell us what is going on today. We reply within one business day with a free initial review.

Talk to an engineer Services