SteppeFin.

The challenge.

Financial users want answers in plain language — what happened to my balance, why did this payment fail, what are my options — but a finance product cannot afford confident nonsense. The assistant had to be grounded in the user's actual data, stream responses, and hand sensitive actions to deterministic flows.

What we built.

We built a Next.js application with the Vercel AI SDK and OpenAI models, where the assistant uses tool calls to read account data and propose actions, and every proposed action goes through a conventional confirmation flow. Responses stream token by token, cite the data they used and degrade gracefully when the model is uncertain.

• Streaming chat assistants embedded in account and transaction views
• Tool calling over typed APIs for balances, transactions and products
• Guardrails: structured outputs, scope limits and human confirmation for actions
• Conversation analytics for quality monitoring and prompt iteration
• Responsive UI with accessibility and keyboard support