Client
High-volume messaging team
Service
Implement
Status
Shipped
Stack
Persona generation · sequential testing · two-model design
A campaign team's synthetic focus group.
AI personas shaped to the team's customers test message variants before a real send. Sequential early-stopping calls the winner, and a synthesis explains why.
The brief
A team sending high-volume customer messages wanted to know which version would land, without spending a real send to find out. Testing on live customers costs reach, and a bad variant costs trust.
What we built
A tool that spins up a synthetic audience to match their customers: AI-generated personas shaped to the demographics they care about, at whatever scale a test needs. The message variants run past that audience, each persona reacts and votes, and results stream in live.
Variants2–4 messages
Personasto your audience
React & voteper persona
Testearly-stop
Winner + why
Two models · a fast model runs the votes · a stronger model writes the analysis
What makes it hold up
The test is sequential, the same early-stopping idea clinical trials use, so it stops the moment one variant is conclusively ahead instead of polling everyone. Results break down by the customer segments the team cares about, so a variant that wins overall but loses a key group gets caught. And a synthesis step explains why the winner won, so the team learns something even when the result surprises them.
Where it landed
Message testing moved ahead of the send instead of after it. The team runs variants past hundreds of personas, gets a winner with the reasoning attached, and spends real sends on messages that already survived an audience.
AudienceHundredsof personas, shaped to your customers
Variants2–4tested at once, with early-stop
Design2 modelsfast eval + strong synthesis
Read another case
All work → 
