Tatiana: A Corporate AI Assistant
Meet Tatiana — our corporate AI secretary.
Tatiana was built using state-of-the-art artificial intelligence technology. Her main job: to become a reliable assistant for employees, streamline workflows, and meaningfully boost team efficiency. With Tatiana, routine tasks get done faster, less time is spent hunting for information and organizing to-dos, and people can focus on the strategic parts of their work.
In Q2 2025, L5's leadership decided to pursue an AI-first strategic approach to the company's development — an approach where AI isn't just a supporting tool, but becomes core to the business.
By Q3 of that year, the Tatiana AI assistant was chosen as the first tool to roll out under this strategy. Her exact feature set wasn't clearly defined at first, but the core idea — task management and video call note-taking — was settled on as the starting point for development.
Tatiana's architecture handles intelligent processing of incoming data, working seamlessly with audio, text messages, Telegram callback responses, and chat conversations.
Classifying Incoming Data
- The classifier receives events from SMTP — the standard network protocol for sending and routing email — and from the Telegram API (messages, callback chats), then determines the content type: audio, text, callback, or conversation.
Launching Cloud Functions and Audio Recognition
- When audio is detected, the classifier sends the file to Yandex Object Storage and triggers an event that launches a Yandex Cloud Function.
- The Yandex Cloud Function handles the integration and launches SpeechKit for asynchronous or streaming audio recognition. Audio recordings are converted to text; files up to four hours long and multi-channel analysis are both supported.
Managing AI Agents and Preserving Context
- AI agents running different prompts work with a local chat database to store short-term context within a conversation.
- Yandex DB is used to enrich her knowledge and provide access to the corporate knowledge base.
- Yandex GPT-5 is used to process text and decide which tool to call — Wiki, Tracker, or SpeechKit.
Service Integrations
- A Yandex Cloud Function handles preliminary transformation of incoming data — for example, text normalization, tagging, or task extraction.
- API integrations allow tasks to be created automatically, the Wiki to be updated, and audio to be transcribed.
Implementing the Agent's "Physical" Presence
- To emulate Tatiana's participation in meetings — on Telemost, for example — a Selenium WebDriver instance is launched. This lets her act through a web interface and join a video call as a virtual participant.
Key Architectural Advantages
- A whitelist of authorized users has been compiled. The bot checks Telegram before every response. Unauthorized users are denied with an "Access Denied" message.
- The task list from Tracker is protected. An unauthorized user cannot retrieve the list of tasks, projects, or goals. The
/tasksand/projectscommands are available only to authorized users. - Corporate data is protected. Access to the Wiki, Drive, meeting minutes, reports, and voice message transcripts is available only to authorized employees.
- Authorization is checked on every request. Every command checks authorization, not just at
/start. Tested: repeat requests also require authorization. - Whitelist storage. The list of authorized users is stored in a secure location (YDB / Lockbox / environment variables) rather than in plain code.
- User list management. Users can be added or removed without a rebuild — via an admin command or configuration.
- Each user's conversation history is isolated. One user cannot see another user's conversations.
