CASE STUDY · /ai-assistant-chatbot

Paybito Whizzo — Domain-Restricted AI Chatbot

A hardened single-domain AI chatbot backend — refuses off-topic and code-gen prompts by design, grounded in a curated JSON knowledge base.

ROLE

Solo Lead Engineer · Backend + Prompt Architecture

TIMELINE

2025

TEAM

1 (solo)

STATUS

Production

Overview

Paybito Whizzo is a focused, single-tenant AI chatbot backend wired to a curated JSON knowledge base. It answers Paybito-related questions strictly — and refuses everything else (jailbreak attempts, off-topic chat, code generation requests).

The discipline is in the prompt and middleware layers, not in the model. Rate limits, validation, structured logging, and a swappable services layer make it operate-able without surprises.

The problem

Generic LLM chatbots happily answer anything — including jailbreaks and off-topic prompts. For a regulated/financial product surface, the chatbot has to be aggressively scoped.

What I built

Hardened system prompt: refusal rules for off-topic, code-gen, and prompt-injection patterns
Knowledge base injection: structured JSON read by dataService → context for every completion
Rate limits: 50/15min on /api/ask, 100/15min on read endpoints — separate buckets
Validation + sanitization on inbound messages; Winston structured logging on every call
Health + data-verification endpoints for ops monitoring
Services / controllers / middleware split — model layer swappable without touching business logic

Tech stack

Node.jsExpressOpenAI ChatGPTWinstonExpress Rate LimitJoi-style validation

Outcome

Production-live as Paybito's public-facing AI assistant. Refusal-rate on off-topic prompts effectively 100% in tested cases. Clean swap-path to upgrade the model layer when GPT-4o-mini → newer Claude/GPT releases ship.

Want this in your stack? Let's talk.

Email Athar Back to all projects