AI agents in a sovereign environment
Data processing and task execution in infrastructure controlled by the customer.
← All solutions
BTG / AI SOLUTIONS
AI agents and digital assistants
AI development and implementation for businesses and public-sector organizations: from a conversational assistant for employees to a fully autonomous solution in a sovereign environment on your own compute infrastructure.
01
How BTG approaches AI
BTG develops AI agents, generative solutions and digital assistants that work with corporate knowledge, documents and information systems. We design the architecture around each customer’s objectives.
Our solutions help people find information, prepare documents, analyze data and carry out approved operations in business systems. Depending on requirements, we use open-weight models that are available without license fees under their terms of use, as well as commercial models and vendor platforms, including Elasticsearch.
02
What we build
AI agents for GPT Experience
A conversational interface to the organization’s knowledge and services, using plain language.
Generative AI agents
Drafts of reports, briefs, letters, presentations and other materials based on set templates.
Digital assistants
Text and voice support for employees, customers and operators in day-to-day work.
GPU appliances
A GPU workstation with a deployed AI agent for closed environments.
BTG can combine these areas in one product: the agent finds information in approved sources, drafts an answer, prepares a document and hands the result to an employee for review.
03
AI agents in a sovereign environment
For organizations that must control where information is stored and processed, BTG builds solutions on local servers, in the customer’s own data center or in dedicated infrastructure. The customer controls access to data, compute resources, models and operating rules.
Working with internal knowledge
Regulations, instructions, internal policies, reports and archives become a searchable knowledge base linked to the source documents. The RAG approach first retrieves relevant passages and then uses them to prepare the answer. For figures and statistics, the agent queries structured data and uses verifiable calculations.
Fully closed environment
For an isolated network, the language model, search, document processing, interface and supporting services — including speech recognition and synthesis — are hosted locally. We verify autonomy across the whole component chain: licensing, telemetry and updates.
Example tasks
Searching internal policies and instructions, briefs for management, answers to employee questions, summaries of internal materials and draft reports grounded in corporate sources.
Access and action control
- Corporate authentication and role-based access control.
- Document permissions respected during search and answer generation, including access by department and dataset.
- Logging of requests and actions in line with requirements for sensitive information.
- Restricted set of agent tools; significant actions confirmed by a responsible employee.
- An agreed procedure for updates, backup and administration.
04
Conversational and generative solutions
AI agents for GPT Experience
Users describe a task in plain language and get the result in one interface. The agent can find an instruction, gather information from several sources, check a request status or draft a ticket. Available actions depend on the user’s role and the organization’s rules.
Generative AI agents
Materials prepared from source data and corporate templates: reports and analytical briefs, drafts of letters, memos and minutes, presentation structure and key points, summaries, classification and document version comparison. Calculations run in separate tools, and documents that need approval are reviewed by an employee.
Digital assistants
Text and voice assistants answer recurring questions, explain procedures, help fill in requests and route complex cases to a specialist. They work in Russian and Kazakh, tested on the customer’s terminology, and support HR, IT and training.
05
Models and technology platforms
We choose the stack by solution quality for the specific task, operating conditions and total cost of ownership. The model handles understanding and generation, the search platform handles retrieval, and the agent logic handles the sequence of steps and tool use.
Open-weight models
For on-premises solutions we use models whose weights can be self-hosted. We select the model after testing on the customer’s data: Russian and Kazakh, document handling, answer quality, speed and GPU memory requirements.
Commercial models and platforms
When a task requires vendor support, we include a commercial model with licensing and deployment constraints in mind. External APIs are used only for scenarios that allow data to leave the environment.
Elasticsearch for search and RAG
Full-text, vector and hybrid search over the corporate knowledge base. BTG is an official Elastic partner and integrates the platform with data sources, the model and the agent interface.
Architecture fit for the task
We consider document volume, update frequency, exact and semantic search needs, existing infrastructure and support costs. The stack is validated in a pilot before scaling.
06
GPU appliance for closed environments
A hardware and software appliance: a GPU workstation with a deployed AI agent for working with data locally. Suitable for a pilot, for equipping a department and for isolated networks.
Compute
A workstation with GPU, CPU, memory and storage sized for the model and load.
Local AI environment
The language model, its serving runtime and document-processing models.
Agent and interface
Agreed scenarios, tools, output templates and session management.
Knowledge base
Loading, indexing and search across approved documents, with OCR for scans.
Operations tooling
Accounts, access rights, logs, monitoring and backup.
Commissioning
Configuration, scenario testing, documentation and user training.
The configuration depends on model size, context length, number of concurrent users and target response speed. For heavy load and high availability we design a server or cluster configuration. Performance is confirmed by measurements on the selected hardware.
07
Implementation stages
Assessment
Processes, data sources, roles and constraints. Priority scenarios.
Design
Model, search stack, integrations, environment boundaries and hardware.
Pilot and MVP
A limited feature set on agreed data.
Development
Scenarios, interfaces, templates and connections to business systems.
Acceptance
Quality, access, performance, autonomy and training.
Support
Knowledge base updates, error analysis and new features.
08
How we measure results
- Answer quality on control questions and correct references to sources.
- Data extraction accuracy and adherence to access rights.
- Time to find information and prepare documents; share of tasks completed with the agent.
- For appliances: speed and stability under the target load.
Target values are agreed before the pilot. The customer gets an AI tool tuned to its own data and processes, with clear access and operating rules, and a phased rollout makes it possible to confirm value before expanding investment.
10
Frequently asked questions
Can an AI agent run without internet access?
Yes. In an autonomous configuration the model, search, document processing and interface are hosted locally, and updates are delivered through an agreed package transfer procedure.
Is Kazakh supported?
Yes. We select the model after testing it in Kazakh and Russian on the customer’s terminology and scenarios.
Does open-weight mean free?
Some models have no license fee, but the project still includes hardware, integrations, support and operations.
Where do we start?
With an assessment and a pilot on a limited set of scenarios. Evaluation criteria are agreed in advance, and the rollout scale is chosen based on the results.
BTG / AI / KAZAKHSTAN
Discuss an AI solution for your organization
Tell us which tasks you want to automate, what data the agent will work with and whether you need a fully closed environment. We will propose the right format: a digital assistant, a generative agent, a sovereign solution or a GPU appliance.