-
Edge-first CNN (EfficientNet-B3) quantized to 8-bit INT → 3.2 MB on-device model.
-
Federated learning loop every 14 days → new species weights without centralising user photos.
-
Safety layer is a rules-engine (Python + SymPy) that cross-checks AI confidence against a toxin database—think seat-belt for tensors.
Full Story
When we open-sourced the MycoGuard dataset two years ago, we had 41 k images across 640 species—nice for academia, scary for production. The break-through came when we re-labelled 22 % of the “edible” class with look-alike tags, forcing the network to learn intra-class variance instead of brute-label memorisation. After INT8 quantisation we still hit 94.1 % top-1 accuracy on our hold-out set while fitting inside the 4 MB Play Store asset limit.
When we open-sourced the MycoGuard dataset two years ago, we had 41 k images across 640 species—nice for academia, scary for production. The break-through came when we re-labelled 22 % of the “edible” class with look-alike tags, forcing the network to learn intra-class variance instead of brute-label memorisation. After INT8 quantisation we still hit 94.1 % top-1 accuracy on our hold-out set while fitting inside the 4 MB Play Store asset limit.
UX Trade-offs We Lost Sleep Over
-
Single-tap vs. multi-angle flow: A/B showed 68 % of users abandon after second photo request, so we made multi-angle optional but surfaced accuracy gain (≈+6 %) with a tiny animated graph—engagement rose 27 %.
-
Red-screen panic: Pure red raised heart-rate in wearable tests; final palette is crimson gradient with a downward chevron—conveys danger without triggering fight-or-flight.
Ethics & Liability
We carry $2 M E&O insurance, but code-wise we hard-fail any prediction below 85 % confidence and force human expert escalation—implemented as a gRPC call to a partner mycology clinic. Average callback time: 11 min, 4 s.
We carry $2 M E&O insurance, but code-wise we hard-fail any prediction below 85 % confidence and force human expert escalation—implemented as a gRPC call to a partner mycology clinic. Average callback time: 11 min, 4 s.
Result
1.3 M active foragers, zero liability claims, 47 documented “saves” self-reported via in-app survey. Edge inference + safety seat-belt = reason you can download and still stay alive.
Grab the repo SDK or consumer build → mushroomcheck.com
本文来自投稿，不代表独立开发前线立场，如若转载，请注明出处：https://91wink.com/building-a-death-cap-detector-the-engineering-diary-behind-mushroomcheck-architectural-tldr/
