AKRIKHIN launched a corporate AI platform within Yandex Cloud in just three weeks

What this is about
AKRIKHIN

Building an enterprise AI agent platform for employees
At the project’s inception, most AKRIKHIN employees did not use AI on a regular basis. A few specialists individually subscribed to various domestic or international LLM services. The company aimed to build internal AI tools that would provide employees with secure access to advanced language models, while establishing a unified platform for enterprise AI agents.
It was critical to minimize the risks of sensitive information leaks, adhere to internal security guidelines, and maintain compliance with national data privacy regulations. Concurrently, the company sought to avoid significant expenditures on integrating and maintaining commercial solutions from external vendors.
AKRIKHIN evaluated multiple options:
- Ready-to-go enterprise solutions for executing LLM workflows.
- Managed service offerings by external providers.
- Proprietary development of an in-house web service.
The selection process prioritized minimizing time-to-market, optimizing project costs, and ensuring infrastructure scalability under heavy workloads. Key criteria included secure data processing, API-driven access to both domestic and international language models, and the ability to launch the project without provisioning dedicated GPU servers. Since commercial alternatives demanded substantial upfront capital outlays or expensive monthly subscription fees, the AKRIKHIN team chose proprietary development.
For project implementation, the team opted for Yandex AI Studio
Deployed an AI agent platform powered by Yandex AI Studio services
The AKRIKHIN team worked on the project from April to August 2025, deploying the initial platform prototype within three weeks. The prototype featured a Go-based web application with access restricted solely to employees authenticated via Active Directory. The pilot phase provided access to YandexGPT Pro and YandexGPT Lite within a unified AI chat interface. The infrastructure utilized Yandex Compute Cloud virtual machines and Yandex Managed Service for Kubernetes® clusters to host the platform, Yandex Smart Web Security to protect the web application against DDoS attacks, automated bots, and SQL injections, and Yandex Security Deck to monitor cloud resources and validate environment configurations.
Following the pilot rollout, the team initiated the development and deployment of AI agents tailored for distinct business use cases, which included:
- RAG agents designed to search information and generate responses based on the company knowledge base.
- An agent engineered to extract text from images.
- An agent configured for price parsing in e-commerce workflows.
- An anomaly detection agent that triggers alerts and notifications.
- A personal assistant integrated with Yandex 360.
Additionally, the team utilizes the platform to aggregate workflow automation requests from business units and build new agents.
The AKRIKHIN team developed a multi-agent architecture. The orchestration tool acts as a single gateway: it handles authentication, accepts incoming requests, routes them to specific agents, and manages session history. An AI agent pipeline implements a designated business use case: the agent constructs the context, transmits the request to the language model, and interfaces with analytical databases, online parsers, or other tools.
For most AI agents, AKRIKHIN specialists selected DeepSeek from the Yandex AI Studio model catalog, while employing external LLM services for specific workflows involving the collection of external insights, such as market-wide pharmaceutical pricing.
Every few months, the AKRIKHIN team runs a cost-benefit evaluation across all models available within Yandex AI Studio to select the most suitable configuration options for both chats and agents. Following the release of Alice AI LLM, the team integrated it into AI chats for employees. Additionally, DeepSeek and Qwen became also featured in such chats.
The service incorporates multiple layers of security. AI agent privileges are restricted at the pipeline and action level: each agent accesses only the tools required for its operation, while read and write permissions are strictly segregated. Foundational LLMs do not receive direct access to company systems, file repositories, issue trackers, or APIs. Instead, all operations route through a Tool Gateway in accordance with current OWASP® guidelines: restricting the toolset available to agents, segregating read and write permissions, and explicitly authenticating sensitive operations.
Prior to transmitting data via API, the service anonymizes personal data. The same logic functions within the AI chat interface when uploading files and text queries. Additionally, the team implemented prompt sanitization and verified the platform’s resilience against prompt injection attacks. Internal testing by the security team and a subsequent penetration test confirmed knowledge base isolation, revealing zero critical vulnerabilities with business impact risks.
Internal employee training was structured as a dedicated phase. AKRIKHIN ran a series of webinars and workshops focused on generative AI features, advanced language models, and their business use cases.
Optimized AI platform expenditures and advanced AI adoption culture
AKRIKHIN launched its AI platform without recruiting additional headcount or incurring significant capital expenditures. According to company estimates, the licensing costs for ready-to-go alternatives from external vendors ranged from ₽6M to ₽50M. Leveraging its internal engineering team alongside Yandex Cloud services enabled the company to deploy a pilot platform within just three weeks, bypassing an external integration project that was projected to take 3 to 6 months. Furthermore, a proprietary platform is significantly easier to scale for new business use cases than modifying a third-party out-of-the-box solution.
Additionally, staff completed training and integrated AI utilities into their daily workflows, sparked by emerging initiatives across various business units to build custom agents.
Moving forward, AKRIKHIN is going to expand the number of AI agent business use cases and further scale the platform.
Contents:

