Case study · Builder.ai

Builder Studio: talk your app into being

Natasha, a voice-first AI that lets non-technical people describe an app out loud and get ready-to-build requirements back.

Problem and role: Turning an idea into app requirements was full of jargon and usually needed an expert. As Design Manager I led the squad of product designers who designed Natasha.
Team and scale: 4+ of the 8 designers I managed worked on this flow. Jul 2024 to Sep 2024. SaaS, no-code and low-code. Live.
The idea: Replace forms and jargon with a guided conversation, and let people speak instead of type.

Natasha's voice-first entry: Let's Talk, Let's Build, with the Natasha orb between Chat and Talk — fig. 1 Speak your way to a fully built app.

Outcome

product designers in the squad I led

Live

shipped in 2024

User feedback showed that speaking felt more natural than typing, so we built voice-first.

Speaking felt more natural than typing.
What users told us, and what changed the product

Problem

Typing out an app idea is hard

Non-technical users struggled to translate their ideas into app requirements. The process was complex, full of technical jargon, and often needed an expert to step in, which slowed the whole journey down.

Our research surfaced six pain points with the chat-based flow.

Incomplete information

People often left details out, which weakened the Buildcard they ended up with.

No natural flow

Typing doesn't recreate the back and forth of a real conversation, so detail gets lost.

Typing errors

Typos led the AI to misread input, which caused frustration and wrong responses.

Time consuming

Typing long, detailed messages is slow and tedious.

Monotonous

Text-only chat felt repetitive and impersonal, without the warmth of a voice.

Context switching

In text it is harder to multitask or jump between topics quickly.

Demo

See it in action

A walkthrough of the voice-first flow, from the first spoken idea to a Buildcard.

fig. 2 The voice-first flow, end to end.

Research

What the existing experience got wrong

The old flow felt fragmented. Instructions were hard to follow, people jumped between steps and documents, and nothing guided their decisions. Research showed that upload was barely used, voice only worked as dictation, and most people saw the assistant as a support chat rather than AI.

Annotated base landing page and refine idea page, with research notes: upload not used, audio sent as a chat message, seen as a support panel, voice limited to dictation — fig. 3 Research notes on the base landing and refine idea pages.

Solution

Two shifts

We set out to make giving requirements feel like talking to a person, and to make what we captured more complete.

Illustration of a speech bubble, a microphone and a document for voice-driven requirement capture — fig. 4 Natural language conversations for app requirements.

Illustration of a document, a chart and check marks for complete and accurate requirements — fig. 5 Guided prompts for complete, accurate Buildcards.

Decision 01 · Benchmark

Learn from the best conversations

We benchmarked Gemini, ChatGPT and Pi.ai to see how each handles flow, engagement and moving between contexts.

Gemini's strength with long-form conversation
ChatGPT's responsiveness in real time
Pi.ai's one-conversation-at-a-time clarity

Voice interfaces from ChatGPT, Pi.ai and Gemini used as inspiration — fig. 6 Inspirations: ChatGPT, Pi.ai and Gemini.

Decision 02 · Input

Make voice the obvious choice

We explored many ways to bring talking, chatting and attaching files into one bottom bar, and gave voice the lead role, from a single Talk to Natasha button to a live waveform while she listens.

Explorations of the bottom bar: chat, talk and files buttons, Talk to Natasha, and recording bars with waveforms — fig. 7 Ideation around the bottom bar.

Decision 03 · Presence

Give Natasha a presence, and a voice

We designed Natasha's states to mirror a real conversation, so people always know what she's doing: a pulse or waveform while she listens, a gentle loading cue while she thinks, and smooth, expressive motion while she talks.

We tested voices on Veed.io and picked one with the right mix of warmth and professionalism. Small human phrases like "That's a great idea" and "You're welcome" made her sound less mechanical.

Natasha's visual identity, an orb, and explorations of her idle, listening and speaking states — fig. 8 Natasha's identity and interactive states.

In action

The platform, end to end

Key interactions and navigation in the final experience.

fig. 9a

Seamless move to voice

The homepage turns into a voice-first interface, with Natasha listening, thinking and speaking from the first moment.

fig. 9b

Explore your way

Talk to Natasha or browse freely. The interface shifts focus based on what you do.

fig. 9c

Interactive cards

Natasha shows short answers and actions as cards right inside the chat, so decisions stay quick.

fig. 9d

From conversation to exploration

A smooth handoff from the chat to a detailed product page gives more context without breaking the flow.

Watch the full 3-minute walkthrough →

Launch

Live in 2024

Builder Studio shipped with Natasha as its voice-first front door, so non-technical users can talk an idea through and get ready-to-build requirements back, without expert help.

fig. 10 The launch highlight reel.

Reflection

What I learned

Voice feels effortless

We assumed people would prefer typing. Feedback showed that speaking felt more natural, so we went voice-first.

Design with mental models

People expect voice to work like a human conversation. Following familiar patterns from ChatGPT and Gemini made Natasha easier to use.

Collaboration makes it better

This project needed far more co-creation than usual. Brainstorming and constant feedback across the squad gave us richer ideas and a more cohesive experience.

Next case study

Agentic AI