Celebrity Like Me
A face-matching entertainment app where the engineering problem was running inference cheaply and handling biometric data responsibly.
- Year
- 2024
- Services
- Mobile appsAI & data
The problem
Face matching is computationally expensive and legally sensitive. Run inference in the cloud and you are transmitting and storing biometric data, with everything that implies. Run it on the device and you have to fit a useful model into a mobile power and memory budget.
What we built
On-device inference using a quantised model, so a user's photograph never leaves their phone. The comparison runs locally against an embedded reference set; what the server sees is a result, never a face.
- Quantised on-device model tuned for mobile inference
- No biometric data transmitted or retained
- Graceful degradation on older, slower hardware
- Plain-language consent before camera access
- Shareable result cards containing no source imagery
The responsible-by-default choice
Keeping inference on the device cost more engineering effort than calling a hosted API. It also meant the honest answer to what happens to your photo was nothing - it stays on your phone. For anything involving biometrics that is worth the extra work, and it removes an entire category of breach risk.
