AI inference inside your perimeter from pilot to industrial operation
We deploy and support large language model inference on your hardware or in a dedicated perimeter. Data does not leave the boundary. We hand over documentation, diagrams and runbooks in Russian (English on request).
We work under contract. The full document set is agreed before signing.
| Site | industrial plant |
|---|---|
| Detection range, 2.4 GHz | 3.1 km |
| Jamming range, 2.4 GHz | 1.8 km |
| Set-up time | 4 min |
| Jamming sector | 90° × 2 |
| False alerts per 8 h | 2 |
Conditions: line of sight, 12 m mounting height, Mavic 3 class airframe, 68 % humidity
Who we work with
Customers whose task is already defined but who lack the resource to take it into operation.
System integrators
You have the customer and the responsibility for delivery, but lack compute capacity and experience with open models. We deploy under your brand.
Enterprises with in-house AI
The model runs in a test environment but never reaches industrial operation: no process, no monitoring, nobody accountable for incidents.
Customers with infosec requirements
The scenario contains personal data or falls under internal information-security rules — computation must run inside the perimeter.
AI service providers
You need peak capacity for projects or load testing — rent accelerators by actual runtime, with no capital spend on hardware.
The infrastructure baseline enterprises bring to AI
The figures below come from published industry research for 2026. We cite the sample so you can verify them yourself.
The pilot is not the problem. Industrial operation is.
Downloading a model takes an evening. Turning it into a service you can trust takes a quarter.Sources: published industry research, 2026. Methodology and sample available on request.
Which delivery format fits you
The formats differ in one thing: where the computation physically runs. That depends on whether the scenario handles personal data, not on the budget.
The same inference and the same models as in the cloud — inside your perimeter. The on-premise version is not a cut-down one: the location changes, not the capability.
Deployment in your perimeter
The solution is installed on your hardware or on dedicated servers inside your perimeter. Requests, weights and logs never leave the boundary.
- Datainside your perimeter
- Accessfrom the internal network
- Billingproject + support
Dedicated capacity
Capacity under our management: API access from your network; infrastructure, updates and monitoring are on us.
- Dataoutside your perimeter
- Accessvia API
- Billingsubscription per seat
Peak capacity for projects
Additional accelerators for the duration of a peak: pilots, load tests, seasonal projects. Billed by actual runtime.
- Dataoutside your perimeter
- Accessvia dedicated nodes
- Billingper hour of runtime
It starts with one question
It is decided by whether the scenario handles personal data. That determines where the computation may physically run.
| Personal data in the scenario | Delivery format | Where processing runs |
|---|---|---|
| Yes | Deployment in the customer perimeter | Inside your perimeter, on your hardware |
| No | Dedicated capacity under our management | Outside your perimeter, API access |
| No, peak load needed | Peak capacity for projects | Outside your perimeter, billed per hour |
If you are unsure which category the scenario falls into, we start by mapping the data flows. That takes one conversation and is not chargeable.
What exactly gets deployed
An open stack with no licence servers. The composition is matched to the model and the budget; below are the baseline node parameters.
| GPUs per node | 1 to 8 |
| Accelerator type | matched to the model, context length and project budget |
| Models | open models with open weights (Qwen, DeepSeek, GLM class) |
| Interface | OpenAI-compatible API |
| Inference layer | vLLM: continuous batching, quantisation, long context |
| Gateway and routing | LiteLLM: single entry point, rate limits, usage accounting, logging |
| Monitoring | Langfuse: call tracing, cost and answer quality — inside your perimeter |
| Network | isolated segment; no outbound connections to external addresses required |
| Integrations | knowledge-base search (RAG), EDMS, 1C, CRM — via API |
| Fine-tuning | on your data, on the same hardware, with no dataset leaving the site |
The final configuration is fixed in the deployment plan once data flows and load requirements are defined.
Your customer, our infrastructure
You hold the customer and the responsibility for delivery. We hold the capacity and the experience of deploying open models. We cover the technical part; the customer relationship stays yours.
- Deployment under your brand
- Second-line support for your team
- Pricing built around the customer and their load
- Responsibility split agreed at acceptance
- Peak capacity for your projects
What happens at each stage
You know in advance what each stage delivers and on what basis the decision to move to the next one is taken.
Defining the boundaries
We work through the scenarios and data flows: what data is processed, whether any of it is personal, where computation is physically permitted.
Solution and estimate
Deployment plan, hardware requirements, model selection, cost estimate. Every assumption in the estimate is stated explicitly.
Running the pilot
We deploy a bounded perimeter and measure latency and throughput on your workload.
Deploy and support
Industrial rollout, handover of documentation and runbooks to your team, ongoing support under contract.
What stays with you after the project
The solution does not require a permanent connection to the supplier. That can be verified at acceptance.
Foreign trade contract: documents agreed before signing
The contract is concluded with a company registered in Hong Kong. We provide incorporation documents and bank details for review by your legal team and bank. Document set: contract, invoice, acceptance act.
Operation without the supplier
No licence server, no mandatory key renewal, no calls to external APIs. Even if our involvement stops entirely, the system keeps running on your hardware.
Documentation in Russian
We hand over deployment diagrams and runbooks in Russian. Your team, or any other contractor, can maintain the solution independently — the model can be replaced without rewriting the application.
Verifiable at acceptance
Deployment runs in an isolated segment with no internet access. The absence of outbound connections to external addresses is verified by your security team at acceptance.
We do not issue legal opinions.
We hand over the technical documentation, and your legal team draws its own conclusions from it. We help describe the personal-data processing architecture for your Roskomnadzor notification and for the impact assessment.How the estimate is built
There is no public price list: cost depends on configuration and operating model. Below is what the estimate consists of.
Deployment in your perimeter
Priced by GPU count, model composition and scope of work: assessment, installation, integrations, acceptance. The estimate is issued before the contract is signed, with assumptions stated.
Dedicated capacity
A subscription by seats and load level. Covers infrastructure, updates, monitoring and support.
Peak capacity
Billed per hour of actual accelerator runtime. Pilot and test perimeters are priced separately from production load.
| Configuration | accelerator count, memory size, network and storage requirements |
| Load | number of scenarios, request intensity, context length, latency requirements |
| Scope of deployment work | assessment, installation, integrations (EDMS, 1C, CRM), load testing, acceptance |
| Term and payment model | project or subscription, contract term, support scope |
We send the estimate once the inputs are in, with assumptions stated. Pilot perimeter terms are agreed separately — on your scenario and your data.
Data does not leave the perimeter.
The model, the weights and the logs stay on your hardware. Cutting the external route does not affect operation.
What customers usually ask
Does data leave our perimeter?
It depends on the delivery format, and that is decided before work starts. If the scenario handles personal data, we deploy inside your perimeter; processing runs there and no cross-border transfer takes place.
If there is no personal data, dedicated capacity under our management with API access is an option. The split is recorded in the data-flow assessment.
What changed in the regulation?
Since 1 July 2025, processing personal data of Russian citizens using databases located outside Russia is prohibited. Breaching the localisation requirement carries a fine of up to RUB 6 million, and up to RUB 18 million for a repeat breach.
The practical conclusion: if your scenario handles personal data, the compute must sit inside your perimeter. If it does not, there is no restriction on location.
Which legal entity signs the contract?
A company registered in Hong Kong. For your finance team this is a foreign trade contract for the import of services: invoice and acceptance act, with VAT as a rule withheld by you as tax agent and then reclaimed. Above RUB 3 million of obligations the contract must be registered with your authorised bank (a unique contract number, UNK) — that is done on your side and we supply the document set for it.
Incorporation documents and bank details are provided at counterparty-verification stage. The above is the general procedure: your finance team confirms the final scheme and we work through any open points with them before signing.
How are payments made?
Currency, route and settlement procedure are agreed before the contract is signed: rouble and yuan payments, and settlement in digital assets (USDT). We choose the scheme around your finance team's and your bank's requirements and fix it in the contract, so the question does not resurface at acceptance.
Can we use our existing hardware?
Yes. If you already have GPU nodes, we deploy on them — there is no need to buy new equipment. During assessment we check whether memory and bandwidth are sufficient for the target model and offer options: quantisation, a different model, or extra capacity only for peak load.
How long does deployment take?
It depends on the hardware available and the model requirements. A work plan with checkpoints is fixed before the contract is signed; the pilot starts once the deployment plan is agreed.
Which models can be deployed?
Open models with open weights that permit deployment on customer infrastructure. We select the set against the task and quality requirements. An OpenAI-compatible interface is used, so the model can be swapped without rewriting the application.
How do we know answer quality is sufficient?
During the pilot we run your scenarios on the selected model and compare results against your acceptance criteria. A quality and latency report is delivered together with the load-test protocol — the decision to go into production is based on that report.
We are an integrator. Can we work under our own brand?
Yes. We deploy white-label, provide second-line support, and do not approach your customer directly. The terms are fixed in the partner agreement.
What about support after go-live?
Support is a separate agreement: request procedure, communication channels, escalation path. Parameters are agreed around your operating model. Documentation and support are in Russian.
Do we need to notify Roskomnadzor?
The personal-data processing notification is filed by the operator — that is, your organisation. We help describe the processing architecture for the notification and for the impact assessment. The legal opinion is issued by your legal team.
Let us discuss your task
Describe the scenario and the constraints — we will come back with an assessment of whether the format fits, and a work plan.
batony@agent.qq.com