Thato Dutch

Selected Work

Agents and content systems that know what they are, what they’re for, and how to talk to people.

01 —Visit Orlando x Tripadvisor
Client work

Orlando is synonymous with theme parks. Tripadvisor needed to show travellers there was more, local culture, food, neighbourhoods, without replacing what already worked.

Role
Conversation designer
Skills
Persona, conversation architecture, content strategy
Platform
Amazon Alexa, Fire TV

What had to hold

Two voices, a local guide and verified reviews, can confuse as easily as convince. A local persona can tip into performance just as easily as it can build trust.

100k sessions · 3-min average, 4.3 interactions per session · Mercury Award 2022

View case study →
02 —The Fabric Matchmaker x Selvedge & Bolts
Client work

Selvedge & Bolts was closing its physical shop and moving fully online. The owner wanted to do something with AI. The website wasn’t doing the job the shop floor used to do.

Client
Selvedge & Bolts
Role
Conversation designer
Skills
Brand voice, guardrails, live prototype

What broke and what fixed it

v2 asked two questions at once and overwhelmed people. v3 overcorrected into a rigid script. v4 handled a near-miss backwards, leading with what we didn’t have instead of what we did. By v5, every tester said they felt confident enough to buy.

5 rounds of testing · all experience levels, confidence to buy · awaiting relaunch for conversion data

View case study →
03 —Edwin
Personal project

Most AI assistants suffer from a terminal lack of character. I wanted to build the opposite: a trivia host who’s comfortably himself, contemptuous, but never cruel.

Role
Sole writer, conversation designer
Modality
Speech-to-speech, ElevenLabs + Claude
Skills
Character design, voice casting, guardrails

What broke and what fixed it

Early versions didn’t sound like the voice I’d auditioned, the model’s own cadence bled through, and Edwin came out terser than dry. Testers didn’t notice, which told me the writing was carrying more than the voice was. The fix was restraint, reactions held to one sentence, openings locked to a rotation, questions hardcoded. Take away the room to perform and the character is what’s left.

12 testers, 0 drop-offs · 7 replayed a second round unprompted · all 12 said they’d return

View case study →