Skip to main content

AI-Driven Customer Support Chatbot

A multilingual support assistant that answers from approved content, and hands off to a human when the evidence is weak.

QBurst Technologies · 2021–2026

  • Python
  • TensorFlow
  • FastAPI
  • ReactJS
  • Vertex AI
  • NLP
  • Chatbot
  • ML/AI
improvement in user engagement from the multilingual chatbot and e-commerce work at QBurst
~35%
improvement in user engagement from the multilingual chatbot and e-commerce work at QBurst

01 — Problem

The problem

Support teams answer the same handful of questions (order status, refunds, password resets) again and again, across channels and languages.

A generic generative model answers quickly but can confidently invent policy. The goal was fast answers that are also traceable to an approved source.

02 — Role

My role

  • Led the solution design: intent taxonomy, retrieval pipeline and escalation rules.
  • Built the FastAPI orchestration service and the Vertex AI / TensorFlow model integration.
  • Shipped the React chat widget and the feedback loop that fed misses back into the knowledge base.

03 — Architecture

How it fits together

  1. 01 Channels

    • Web chat widget
    • Messaging apps

  2. 02 Orchestrator

    • Language detection
    • Intent classifier
    • Policy & auth checks

  3. 03 Knowledge

    • Hybrid search
    • Approved help content
    • Order & account APIs

  4. 04 Response

    • Grounded answer + citation
    • Human handoff with transcript
The orchestrator decides between an FAQ answer, retrieval, a tool call or a handoff before any text is generated.

Key decisions

  • Intents first, generation second. Account-specific questions always go through an authenticated tool call.
  • The bot refuses when retrieval confidence is low, rather than guessing.
  • PII is redacted before anything reaches logs or analytics.
  • Thumbs-up/down and resolution tags feed a weekly review of unanswered questions.

04 — Results

Results

  • ~35% improvement in user engagement from the multilingual chatbot and e-commerce work at QBurst
  • Routine questions resolved without an agent, with full transcripts when a human takes over

05 — Retrospective

What I’d do differently

Build an evaluation set of real customer questions before choosing a model, so every model change is judged on data rather than anecdotes.

Treat the knowledge base as a product with an owner. A grounded bot is only as accurate as the help content behind it.

Related articleModern Customer Care with NLP and Vertex AI

Working on something similar?

I’m happy to talk through architecture, trade-offs or a role.

Get in touch