AI Conversational Assistant

The problem

Wiring a chat box to a completions endpoint takes an afternoon. Building an AI feature that a team can still maintain a year later is a different exercise, and most of the difference is architectural rather than model-related.

What I built

A React chat interface against a FastAPI backend handling OpenAI API calls, prompt construction, and conversation context and session management. Response streaming for low-latency replies, so the interface starts responding while the model is still working.

The decision that mattered: prompt logic and conversation state stay fully decoupled from the UI.

Why that decision

Prompts change constantly — every week, sometimes every day. Interfaces shouldn't. If your prompt construction lives in your components, every prompt iteration becomes a frontend deploy and every A/B test becomes a merge conflict.

Keeping them separate is also what makes it possible to wire custom agents into an existing product later, rather than bolting a chatbot onto the side of one. The same service layer that answers a chat message can answer a tool call.

Stack

React · Python (FastAPI) · OpenAI API