Case study · Australia · Wellness app
Catchfire AI×
zkAGIWandrlust: proven at scale before launch
Another firm left Wandrlust with a slow app that gave mixed answers. We rebuilt the AI, tested it at scale, and made location privacy the default.
- Client
- Wandrlust, Australia
- Partner
- Catchfire AI, UK
- Delivered
- AI guide, scale test, privacy
- Status
- Live
The story in 60 seconds. Captions are on the video.
The problem
- Slow: answers took more than two seconds.
- Inconsistent: camera, voice and map each gave a different answer.
- Untested and risky: nobody knew what would break at scale, and exact locations were stored.
What we did
- One AI guide that sees, listens and knows the area.
- A simulated million users before launch, to find what breaks first.
- Locations blurred to about 100 metres and deleted automatically.
The result
100%
of sessions passed with 25 users at once
1M
users simulated before launch
45% → 55%
retention the simulation says faster voice can unlock
~100 m
location blur, by default
Only with zkAGI
- Scale tested before launch, not discovered in an outage.
- Privacy built into the platform, not added later.
- An engineer who stays until the numbers move.
See it live
Links
wandrlust.ioWandrLust on XField guideLinkedInzkAGI report, Dec 2025
Under the hood, for technical readers
- Vision and voice on the fast, zero-retention lane; planning on the mixture-of-agents lane.
- Mirofish simulation: 1,000 Monte Carlo runs across four user personas, 95% confidence.
- Redis caching and an attribution pipeline added from the simulation's findings.
Launching something people will rely on?
We will test it at scale before your users do.
Talk to an engineerMore case studies