Inc. covered ESTESIS products — and how we build infrastructure for enterprise AI
Inc. Russia and Cloud.ru ran a special project on seven companies building AI products on a cloud platform, and one chapter is about us: three tools for working with large language models and how, together, they save hours of work and money spent on AI.
What they wrote
Three tools for working with large language models
ESTESIS entered the market in 2019 as a research lab in machine learning and voice biometrics. Later the team focused on end-to-end AI services.
Our own product line grew out of our own needs: build AI agents faster, keep spending on language models under control and protect sensitive data. That is how LLMBox, LLMGW and LLMGuard came about.
As a Cloud.ru partner we use Evolution ML Inference to launch models quickly and compare them by quality and cost. For a gateway that picks a model for every request, this is a key capability.
Example from the article
A production report in minutes instead of hours
- 01LLMBoxAn AI assistant gathers data from documents and enterprise systems, analyses it and prepares the report.
- 02LLMGatewayAn AI gateway. It provides the model that fits the task by capability and cost.
- 03LLMGuardChecks AI requests and responses so that confidential data does not leak to external services.
The LLMCore product line
Five products that take AI from model selection to production
01 / Model gateway
LLMGateway
One API to 300+ models. Routing, usage accounting and cost control
- Cloud
- llmgw.ru
- Single API
- OpenAI-compatible interface: connect once, work with any model
- Routing
- Each request goes to the model that fits by capability and price
- Usage and limits
- Spending by key and project, budgets and cost control
- Payment in rubles
- No foreign cards or separate contracts with every provider
02 / Data protection
LLMGuard
An AI firewall for LLMs. Checks requests and responses and stops confidential data
- On-Prem
- Cloud
- llmguard.ru
- Requests and responses
- Traffic to models is checked both ways, before and after generation
- Personal data and secrets
- ML detection of Russian personal data, keys, passwords and proprietary code
- Policies as code
- Rules at organisation, project and key level; observe first, then block
- Your own perimeter
- Deployed on the client’s servers or in a closed network
03 / Agent platform
LLMBox
An AI agent builder with tools, a knowledge base and any models
- On-Prem
- Cloud
- Agents for the task
- System prompt, MCP tools, RAG and a knowledge base, no custom development
- Any models
- Connected via LLMGateway: external and local models in one window
- Protected by default
- Agent requests go through LLMGuard
- Role-based access
- Permissions by role and integration with enterprise systems
04 / Security testing
LLMRed
Red teaming for LLM systems. Finds vulnerabilities before launch, not after an incident
- Service
- Report
- Attacks on the model
- Prompt injections, jailbreaks and bypassing restrictions
- Leaks
- Checks whether the system gives away confidential data
- Agents and tools
- Tests of unsafe agent actions in external systems
- Report and rules
- Vulnerabilities found become LLMGuard rules
05 / Quality evaluation
LLMEval
Quality metrics for models and agents. Choose a model by the numbers, not by feel
- Service
- Report
- Tests for the task
- Custom checks and metrics for the client’s domain
- Model comparison
- Quality, speed and cost on the same data
- Regression
- Checks before changing the model, prompt or agent version
- Linked to the gateway
- The best model by evaluation becomes a route in LLMGateway
How it works together
One perimeter from the application to the model
- Applications and agentsLLMBox
- ProtectionLLMGuard
- GatewayLLMGateway
- ModelsOwn and external
LLMEval and LLMRed test the system before launch and on every change
Read also
Research and agentic systems
All articlesAccredited IT company · Russian Ministry of Digital Development registry No. 20142