SHOPIFY CASE STUDY
ChatNT: A Voice-Led Generative-AI Companion for Summer in the Northern Territory
ChatNT is the AI assistant behind Tourism NT’s “Chat with Abbie Chatfield” campaign. Abbie joined as Chief Sensory Officer for Summer in the Northern Territory, so we cloned her voice from her own audio: ChatNT answers summer-travel questions in Abbie’s voice, grounded in trusted content, trained on real data and team feedback, and held strictly to the summer brief.
Project Context
Client
Northern Territory (Tourism NT)
Niche
Generative AI · Voice AI · Travel & Tourism
The Tech Stack
- React.js
- TypeScript
- Tailwind CSS
- Python (FastAPI)
- Node.js
- LangChain
- OpenAI GPT-4.1 / GPT-5
- ElevenLabs
- Pinecone
- MongoDB
"We envisioned a travel companion that inspires visitors to discover the best of the Northern Territory while providing accurate guidance and filtering out unrelated or sensitive topics."
Those demands pull against each other: inspiration wants an assistant that talks freely, accuracy wants it tied to sources, safety wants it to refuse, and the campaign wanted all of it in Abbie Chatfield’s own voice. Reconciling them is what this build is about.
Northern Territory Team
Tourism Northern Territory
HOW WE BUILT IT
A Structured Build Process for a Safe, Accurate Assistant
Delivered in a month by a three-developer team, the project ran across voice cloning, knowledge design, a RAG pipeline, response-model training, guardrails, personalisation, and context management. We agreed the summer scope and safety rules up front, then trained and tuned in fast iterations.
DISCOVERY
A working session to define the summer campaign, the traveller journeys to support, and the hard line between on-brief help and everything Abbie must not answer.
REQUIREMENT ANALYSIS
Split the brief into workstreams: voice clone, knowledge base, RAG retrieval, response-model
training, guardrails, personalisation, and context management.
Knowledge & RAG Design
Structured the trusted Summer-in-the-NT content, chose the chunking and embedding strategy, and
designed the Pinecone retrieval that grounds every answer.
Voice Cloning
Cloned Abbie Chatfield’s voice from a library of her audio using AI voice models, so ChatNT’s answers are spoken in her voice, delivered through ElevenLabs.
Model Training & Feedback
Trained and tuned the response model on curated datasets and folded in the internal team’s testing
feedback, iteration after iteration.
Guardrails & Scope
Built the intent and scope classification, plus re-prompting, that keeps every reply on Summer in the NT and refuses the rest.
Personalisation & Context
Built the itinerary logic and the context-window management that keeps multi-turn planning coherent
across a whole conversation.
QA & Launch
Adversarial guardrail testing, accuracy checks against the knowledge base, and voice and latency tuning before going live on the campaign site.
Conversational Discovery with Abbie
The traveller-facing experience, “Chat with Abbie Chatfield”, turns summer trip planning into a conversation. Suggestion chips and natural language let a visitor start from a vibe (flowing waterfalls, epic wildlife, red-dirt drives), and Abbie takes it from there.
Abbie's Voice, AI-Cloned
Abbie joined the campaign as Chief Sensory Officer for Summer in the NT, so ChatNT had to sound like her. We cloned her voice with AI voice models trained on a library of her audio and delivered it through ElevenLabs, so answers are spoken in Abbie’s own voice, not a generic text-to-speech.
Retrieval-Augmented Grounding
Every answer is grounded. A LangChain retriever pulls the most relevant passages from a Pinecone index of trusted Summer-in-the-NT content, and the model generates its reply from those sources, so ChatNT informs rather than invents.
Trained on Real Data & Team Feedback
Response quality came from training, not luck. We trained and tuned the model on curated datasets and folded in feedback from the internal team’s testing rounds, tightening accuracy on real summer-travel questions with every iteration.
Guardrails & a Strict Summer-in-the-NT Scope
An intent and scope classifier sits in front of every reply. On-brief summer questions get an answer; anything unrelated or sensitive is re-prompted back to summer travel; unsafe input is blocked with a safe fallback. Abbie stays on message.
Personalised Summer Itineraries
The assistant turns trip duration, interests, budget, and season into a real day-by-day summer plan, Darwin and the waterfront, Kakadu and Litchfield, Uluru and the Red Centre, instead of a list of links.
Cloud-Native Architecture
A React and TypeScript front end, a FastAPI and Node backend, an AI layer of LangChain, OpenAI, and the ElevenLabs voice clone, and a data layer of Pinecone and MongoDB, built to scale with monitoring and analytics.
Managed Context & Re-Prompting
Context carries across the whole chat: say “three days this summer” in one message and “I love swimming holes” in the next, and both shape the plan. Re-prompting continually steers the model back to the Summer-in-the-NT brief so it never drifts off scope.
WHAT TO EXPECT FROM US
Transparent Build Process and Clean Delivery for AI Projects
If you are building a generative-AI product and weighing partners, here is what working with our team actually looks like.
A plan before a proposal.
We start by defining what your AI must do, who it serves, and, just as important, what it must refuse. You leave with a delivery plan and a clear domain boundary, not a sales pitch.
Grounded in your data, not guesswork.
We build retrieval over your trusted content so the model answers from sources you control. Accuracy comes from grounding, not from hoping the model gets it right.
Safety and guardrails from day one.
Topic control, refusals, and safe fallbacks are part of the architecture, not a filter bolted on at the end. A public assistant has to be safe before it is clever.
Clean, maintainable code.
We write front-end and backend code any competent engineer can read and extend. No black-box glue that no one can maintain after launch.
A milestone review at every stage.
You review the real system at each stage: knowledge model, retrieval, guardrails, personalisation, final QA. You always know what works and what is left.
Quality that does not depend on one person.
Guardrails are adversarially tested, accuracy is checked against the knowledge base, and latency is measured. Nothing ships because it worked once in a demo.
You own what we build.
At handover you get the full stack: source, prompts, retrieval configuration, infrastructure, and documentation. Your team can run it independently or keep us on; your choice.
A founder-friendly working rhythm.
Fast iterations, honest trade-offs, and a build shaped around your actual goals. You are a partner in the build, not a ticket number.