Case study · AI prototyping
How When I Go got a working AI avatar prototype in two weeks
A voice and avatar experience for one of the hardest conversations a person can have, proven with a working demo before the production build started.
In short
BlueButterfly is building When I Go, an AI-guided end-of-life planning product. Before committing to the production build, the founder needed proof that a voice and avatar experience could guide people through hard, personal decisions and feel human doing it. I designed and built two working prototypes in two weeks with Bolt.new, Supabase, HeyGen, and ChatGPT. The prototype became the benchmark for the production build, and When I Go launched with April, the avatar companion, at the center of the product.
| Client | BlueButterfly (When I Go, whenigo.io) |
|---|---|
| Industry | Consumer, end-of-life planning |
| Role | Product contractor |
| Engagement | Sprint, concept to demo |
| Duration | Two weeks to a production-quality demo |
| Stack | Bolt.new (front end), Supabase (database and APIs), HeyGen (avatar), ChatGPT and the ChatGPT Voice API (conversation) |
| Outcome | Prototype became the benchmark for the production build. When I Go launched with April at the center of the product. |
What was the problem?
A new company set out to reinvent end-of-life planning with AI. The product asks people to make decisions most of us put off for years: what happens to our bodies, our belongings, our words to the people we love. The hard question was whether a voice and avatar experience could carry that weight: guide someone through a long, emotionally sensitive planning form, keep the conversation moving toward completion, and feel warm and trustworthy the whole way.
Nobody wanted to fund a production build on a slide deck. The founder needed something people could try.
What did I build?
Two working prototypes, built from scratch with AI-assisted development tools.
Version 1: April, a conversational video avatar
Built on HeyGen with ChatGPT behind the conversation. April greets the user, explains the planning process, and leads them through the form section by section. The conversation logic steers back toward completing the form whenever the discussion wanders, which matters in a subject this emotional.
Version 2: voice with a live form
A split-screen interface on the ChatGPT Voice API. The user talks on one side; the form fills in on the other side in real time as the conversation happens, so there is no separate data-entry step at the end.
Stack and timing
Bolt.new for the front end, Supabase for the database and APIs, HeyGen for the avatar, ChatGPT for the conversation. A first working version existed within days. The polished, production-quality demo took two weeks.
What changed?
The demo answered the question before the production build started. The founder used the prototype as the benchmark for the production build, and When I Go launched with April at the center of the experience: the live planning dashboard has April available as a companion alongside the plan summary and each section of the plan.
What I learned
- AI prototyping tools have collapsed the time from concept to demo. The gap between a convincing demo and a production-ready product is still wide and still underestimated. Plan for it from the start.
- Emotionally sensitive products need different UX instincts than productivity tools. Pace, tone, and permission to pause matter more than speed.
- A prototype proves the idea; a short strategy document keeps the team building the same thing. Run both, in parallel.
Questions people ask
Can a HeyGen avatar run a guided conversation?
Yes. In this prototype HeyGen provided the video avatar and ChatGPT handled the conversation and the logic that steered it toward completing the planning form. A second version dropped the avatar for voice plus a live form that filled in as the user spoke.
How long does a voice and avatar AI prototype take to build?
Here, a first working version in days and a production-quality demo in two weeks, built by one person with AI-assisted development tools: Bolt.new, Supabase, HeyGen, and ChatGPT. Production readiness is a separate and longer effort.
Have an AI idea that needs proving?
Two weeks and a working demo settle more arguments than a quarter of planning.