Building Janus: An AI Customer Support Helpdesk System Powered by MindsDB

Enterprise support teams face a constant drain on time and resources—triaging repetitive requests, classifying tickets, and responding to the same issues over and over. The result? Slow turnaround times, inconsistent responses, and support agents overloaded with manual work.

Janus was built to eliminate this bottleneck.

Powered by MindsDB AI Agents and a MindsDB-managed Knowledge Base, Janus transforms the traditional helpdesk into an automated, intelligent system. Users can submit tickets and chat with an AI agent in real time, while admins manage insights, classifications, and Knowledge Base updates from a unified dashboard.

The result is a fast, consistent, AI-first helpdesk workflow that automates classification, conversation, and analytics—reducing manual workload and accelerating resolution times.

The Use Case

The goal was to automate the helpdesk lifecycle.

• Users submit tickets and chat with an AI support agent in real time

• The system automatically classifies each ticket by intent, priority, and category

• Admins get a single dashboard to track trends, promote solved tickets into the Knowledge Base, and search historical data

The outcome is faster response cycles, consistent answers, and actionable insights.

Key Features

1. AI-Driven Ticket Classification

Each ticket is processed by a dedicated MindsDB AI Agent that predicts its type, category, tags, and priority. The model learns from past tickets to improve accuracy over time.

Ticket metadata schema:

 {
     "type": "str",
     "category": "str",
     "priority": "str",
     "tag_1": "str",
     "tag_2": "str"
   }

2. AI Chat Support

A second agent handles real-time conversations, providing instant solutions by leveraging the existing Knowledge Base for context.

3. Admin Dashboard

Admins can view analytics, approve solved cases into the Knowledge Base, and filter data by type, priority, or category.

4. Knowledge Base Management

The Knowledge Base (KB) is managed by MindsDB. It stores both content and metadata for retrieval and reasoning.

KB schema:

   content_columns = ["subject", "body", "answer"]
   metadata_columns = ["type", "priority", "category", "tag_1", "tag_2"]

5. Search and Insights

Admins can visualize ticket trends, most common tags, and distribution of categories directly within the dashboard.

How It Works

Below is the high-level workflow of Janus.

Architecture diagram:

• The Ticket Classifier Agent predicts structured metadata in JSON format which can be parsed directly in python using json.loads()

• The Support Agent handles ongoing chat with context from the Knowledge Base

• Both AI Agents are powered by MindsDB which manages the AI and KB layers.

For more information about the architecture of Janus, visit here.

How Janus Was Built.

Tech Stack:

• UI: Streamlit

• AI & KB Layer: MindsDB

• Vector Database: ChromaDB

• LLM Provider: Nebius

• Programming Language: Python

The connection between Streamlit and MindsDB is handled via the MindsDB Python SDK.

AI Setup

MindsDB integrates with external AI providers (having OpenAI compatible APIs) for model inference.

For this project, Nebius was used as the provider endpoint. The LLM used is Qwen/Qwen3-235B-A22B-Thinking-2507, and the embedding model is Qwen/Qwen3-Embedding-8B.

Two MindsDB AI Agents power the workflow:

1. Ticket Classifier – Predicts metadata such as type, category, and priority.

2. Support Agent – Generates AI responses using the Knowledge Base for context.

The Knowledge Base combines a vector database (ChromaDB) with embedding model for contextual retrieval.

“If you’d like to see my rough work including KB creation, AI agent logic, querying, and some files database operations check out the notebook here.”

Installation

1. Clone the github repository

git clone https://github.com/k0msenapati/janus.git
cd janus

2. Download the dataset

# Create data directory if not exists
mkdir -p data

# Download dataset zip via Kaggle API
curl -L -o data/customer-support-tickets.zip \
  https://www.kaggle.com/api/v1/datasets/download/tobiasbueck/multilingual-customer-support-tickets

# Unzip contents to data directory
unzip -o data/customer-support-tickets.zip -d data

# Optional: remove the zip after extraction
rm data/customer-support-tickets.zip

3. Setup project, environment variables and run setup script

uv sync

cp .env.example .env # add your nebius api key from nebius ai cloud to NEBIUS_API_KEY var

uv run setup.py

4. Run Streamlit App

source .venv/bin/activate
streamlit run app.py

Impact

• Reduced response time as users get immediate AI replies

• Consistent ticket classification ensuring better routing and tracking

• Visual insights on ticket categories, priorities, and tag trends

• Easier management of historical tickets through the Knowledge Base

AI acts as the first responder, handling repetitive issues and freeing human agents for complex cases.

Future Enhancements

• Web Search Integration: Adding real-time data retrieval would make the AI agent more dynamic

• Backend Upgrade: Creating a FastAPI backend for production scalability

• Third-Party Integrations: Utilizing MindsDB data integrations such as Jira, Slack, or Gmail for enterprise integration

Final Thoughts

Janus shows what’s possible when modern AI workflows are paired with MindsDB’s agent and Knowledge Base capabilities. With a modular architecture, real-time AI support, automated classification, and fully integrated analytics, it delivers a powerful blueprint for end-to-end helpdesk automation.

And this is just the beginning. As Janus evolves—through web search integration, backend scaling, and deeper enterprise connectors like Jira or Slack—it can grow into a production-grade, multi-channel support system that dramatically reduces operational overhead.

If you’re exploring how to bring AI into your support stack, Janus is a strong example of what an AI-native helpdesk can look like—and how MindsDB enables it. Check out the project on Github.

Demo

Check out Janus the AI Helpdesk system powered by MindsDB in action.