Designing and building a product alone with AI

Overview

Billvox is a WhatsApp tool that turns voice notes into invoices and quotes: a tradie finishes a job and sends a voice note or a quick text - “Done at Murray’s, swapped three downlights, two hours, parts were $85” - confirms the details with a tap, and a GST-compliant invoice or quote arrives by WhatsApp and email. I designed and built it on my own, working with Claude from the first plan through to running code. It started as a learning project. It turned into the clearest view I’ve had of what AI is doing to the way products get made.

Context

When I started out on the web in the early 2000s, getting a product like this in front of people would have needed a small team - a designer, a couple of developers, someone to look after the infrastructure. I lead a large design team in my day job, and I’d been watching AI change what designers spend their time on. I wanted to understand it from the keyboard rather than from a slide.

Billvox is built for tradies and other sole traders - people who spend their day on the job, not at a desk. Many of them lose hours every week to admin - chasing invoices, forgetting to follow up quotes and typing up jobs after a long day on the tools. The software that exists mostly asks them to sit down, open an app and fill in forms, which almost nobody does when they’re dirty and tired at the end of a shift.

Approach

I planned the product in a dedicated Claude project, working through the problem, the scope and the build steps before writing anything, and then built it in VS Code with Claude Code, keeping the project’s context in a single file at the root so each session picked up where the last one left off.

It runs on WhatsApp through Twilio. OpenAI’s transcription turns voice notes into text, Claude turns messages into structured records, and the app generates GST-compliant invoices and quotes as PDFs and sends them back by message and email. Everything is hosted on Railway and Cloudflare, for a few dollars a month.

The interesting part wasn’t the code. It was how much of the work became deciding: what the product was for, what to leave out, what “good” looked like for an invoice someone would actually send, and when something Claude produced was plausible but wrong. The biggest of those decisions was what not to build. Billvox isn’t another job management app - there’s no app at all. It lives in WhatsApp, where tradies already are, sits alongside the systems they already use, like Xero and ServiceM8, and is built around Australian GST and ABN rules from the start. No AI tool made that call for me, and I think it matters more than any of the code.

Outcome

Billvox is now in early testing with tradies. I’m deliberately putting real feedback ahead of new features, which is its own lesson: when building gets this cheap, the discipline is not building ahead of your users.

Reflection

I think the main lesson for design leaders is that the making really has become cheap. What didn’t become cheap was knowing what to make, noticing when it was wrong and holding a clear idea of quality. Those were the bits only I could do, and they were most of the job.

It also made me more sympathetic to the designers I work with. It’s one thing to say AI will change the role, and another to feel how much of your own craft it can now do in an afternoon.