Building a Banking AI Chatbot: When Simple WebSockets Became a Complex Engineering Problem

A Node.js developer working on a production banking AI chatbot shares how what seemed like a straightforward WebSocket integration quickly grew into a multi-layered engineering challenge. During early testing, the team discovered the bot had no mechanism to reject off-topic queries, responding to an IPL cricket question with nonsensical banking fragments pulled from its vector knowledge base. To fix this, the team added an intent detection layer that screened queries before routing them, but the solution introduced a 200–300ms latency penalty on every single request. Attempts to offset this with Redis caching failed because exact-string matching could not recognize that differently worded questions carried the same meaning. The article is the first in a series documenting the real-world engineering decisions and trade-offs encountered while shipping the chatbot to production.