AI consultant widget

Operating as twin platforms, Kadroland and 7eminar offer training and consulting in their respective fields: HR management and accounting. Both platforms built their reputation on providing expert consultations, but human consultants could only handle so many inquiries at once. Availability was restricted to business hours, and peak periods created response delays.

The client saw AI as an opportunity to maintain their expertise while dramatically expanding capacity. Thus, the project became a solution to a practical business problem: how to serve more clients without compromising quality or proportionally increasing costs.

Rather than setting up a pre-packaged chatbot solution, the client team came to Brights with a blank slate. We started from zero: no existing AI infrastructure, no predetermined architecture, just a goal to build something that works at scale.

We didn't jump straight into a full-featured widget. Instead, together with the client, we decided to test the concept with a Telegram messenger bot: a faster, more cost-effective way to validate whether AI could truly match human consultant quality.

This PoC stage proved critical. The client's internal experts tested the bot on 250 questions covering various topics in tax and personnel law, evaluating responses against their professional standards and scoring the AI's performance. When 80% of the answers were rated as excellent, we had proof: the project was actually viable for our client's business.

With validation in hand, we moved to MVP development, designing the solution that now lives on both platforms.

The client's design brief was quite direct: make it visually appealing, modern, and similar to ChatGPT. Not because they wanted to copy, but because their users already understood that interaction model.

We delivered ChatGPT-inspired UI within two weeks across both platforms, iterating through design revisions while keeping the timeline tight. The familiar interface reduced the learning curve, allowing users to engage with the AI consultant without lengthy onboarding.

The widget handles thousands of concurrent users while maintaining 99%+ availability and delivering responses in under 90 seconds. Here's how we built for scale and quality:

Multi-agent system: Different AI agents handle different types of queries, allowing specialized responses rather than one-size-fits-all answers.

RAG (Retrieval-Augmented Generation): The system pulls from vectorized data chunks, ensuring responses draw from the most relevant, up-to-date information in the knowledge base.

Dynamic knowledge updates: New external sources integrate easily, keeping the consultant current as laws and regulations evolve, which is critical for HR and accounting sectors.

Infrastructure ready for growth: Kubernetes deployment on Hetzner infrastructure means the system can scale from hundreds to tens of thousands of active users without architectural changes.

Over the course of this project, we transformed a manual consultation model into an autonomous, scalable AI solution that maintains expert-level quality while dramatically expanding capacity.

Timeline: The MVP was built from scratch and deployed across both platforms in just three months.

Consultant relief: The AI widget handles both routine and complex questions 24/7. Human experts step in when users need additional clarification or personalized advice.

Cost efficiency: The client redirected significant consulting volume without hiring additional staff or expanding team capacity.

Expert-level quality: The client's internal experts evaluated the AI's responses using quality KPIs and rated 80% of them as “excellent” against their professional standards.

Competitive advantage: The multi-agent system we architected delivers superior results compared to the simpler solutions most of our client’s competitors use.

After releasing the AI consultant widgets, our team continued PostMVP development. We've been refining the chatbot and implementing new capabilities on both the frontend and backend.

We're also expanding the multi-agent system to cover more specialized platform topics and have integrated additional LLM models, allowing users to select their preferred model for consultations.