Generative AI Customer Support System

Inspiration

We frequently observed a high volume of recurring questions on open customer support platforms — often involving nearly identical issues. This revealed a clear inefficiency: new tickets were being treated independently, while valuable past resolutions remained unused. We saw strong potential in exploiting historical ticket data and using semantic proximity to group and match new tickets with existing solutions.

What it does

Our system automatically classifies incoming support tickets by topic and urgency, retrieves similar resolved tickets using vector search, and suggests relevant responses to the agent. It can also escalate tickets to human agents when necessary, passing along urgency and context — all in just a few seconds.

How we built it

• We fine-tuned a Gemini model on support ticket data using Vertex AI to predict ticket type and urgency.
• We stored past tickets as embeddings in MongoDB’s vector search, enabling semantic similarity matching.
• The AI reads new tickets, classifies them, retrieves similar tickets, and provides contextual suggestions.
• We implemented function calling to allow AI to reference past tickets and assist in building final responses.
• If the user need a more precise answer, the ticket is escalated to a human with full context attached.

Challenges we ran into

• Getting high-quality, diverse training data for fine-tuning the model.
• Ensuring fast yet accurate vector search across large volumes of past tickets.
• Tuning the urgency prediction model to reflect real business needs.

Accomplishments that we're proud of

• We integrated the entire process into a smooth, end-to-end support flow.
• We created a scalable framework that can evolve with more data and features.
• We produced a fully containerized solution

What we learned

• How to use LLMs and vector databases together to enable smarter automation.
• The importance of combining AI predictions with human-in-the-loop design.
• How to handle ambiguity and variability in real-world customer support data.

What's next for Generative AI Customer Support System

• Multilingual support for international tickets.
• Trend detection across incoming tickets to identify common problems early.
• Full automation of low-priority tickets with agent approval.

Built With

• cloudfare
• docker
• docker-compose
• fastapi
• google-cloud
• google-compute-engine
• javascript
• nginx
• python
• react
• shell
• tailwind
• typescript
• vite