Nova

An AI customer support agent that lives inside your product.

Try the live demo | View the source on GitHub

We built Nova inside a sample neobank dashboard, where it answers customer questions from a knowledge base, looks up account and transaction details, and hands the conversation to a person when a question needs one. The same approach works for any business whose customers ask questions that the team currently answers by hand.

What it does

• Knowledge-grounded answers: Before each reply, Nova pulls the most relevant material from a knowledge base, so its answers stay grounded in real documentation.
• Tool use: it looks up a customer's transactions and account information, and escalates to a person by email when a question falls outside what it can resolve.
• Real-time streaming: responses appear as they are generated, with inline indicators showing when Nova is searching or using a tool.
• Multi-turn conversations: it follows the thread across messages so a customer can ask natural follow-up questions.

How it works

When a customer asks a question, Nova converts it into an embedding and matches it against a knowledge base stored in a vector database. The closest passages go to the model as context, and the model writes its answer, calling tools when it needs account or transaction data. The reply streams back to the customer as it is written. We built it on production foundations, with rate limiting, a fully typed codebase, and a complete test suite.

Tech stack

• Frontend: React and TypeScript
• Backend: FastAPI and Python
• Model: Claude
• Embeddings: OpenAI
• Vector storage: ChromaDB
• Streaming: Server-Sent Events

About this demo

Try it yourself

Try the live demo to chat with Nova on the sample neobank.

View the source on GitHub to see how it is built.

If you want an agent like this inside your own product, let's talk.

Get started