The first private cohort wrapped this week. Forty testers across Zagreb, Split, and Ljubljana, ranging from native Deaf signers to complete beginners, ran the app through real conversations and learning sessions for three weeks.
The good
- Latency held. Mean time from sign to text was 187 ms on iPhone 14 and newer, well under our 200 ms ceiling.
- Fingerspelling broke through. The two-handed HZJ alphabet, our hardest case, hit 92% top-one accuracy in good lighting.
- The learning loop is sticky. Day-three retention was 62%, day-seven was 41%. We did not expect day-seven to clear forty.
The hard
- Low light is still hard. Below about eighty lux, accuracy drops fifteen points. We have a fix in testing.
- Two-hand grammar across signers was inconsistent. Different regions sign the same idea differently, and the model picked sides. We are collecting more balanced data with three new signing partners.
- Sign continuity is harder than fingerspelling. Connecting signs into phrases is where the next quarter goes.
What changed in the app
- Camera exposure now biases for hand regions, not faces.
- The learning path adds DZ and Š disambiguation drills, two letters people consistently confuse.
- We added a small confidence indicator. If the model is not sure, you see it before you act on it.
Cohort two starts mid-May. If you want in, the waitlist is the door.
Keep reading
More notes from the build.