google gemma introduced a tool that turned a budget android phone into an offline sign language translator – for 2.5 million deaf indonesians who don't have reliable internet

kawanisyarat won the cactus prize in the "gemma 4 good" hackathon, beating out 1,600+ submissions. it's a local-first android app bridging indonesian sign and spoken language – entirely offline.

gemma 4 e2b runs on-device. point the camera at sign language – it translates gestures into natural sentences. someone speaks – whisper transcribes, then gemma converts the transcript into simpler text that's easier for deaf users to read. two-way communication, zero connectivity.

the engineering that makes it work: memory-aware model routing. the app loads whisper when it needs speech-to-text, then unloads it so gemma gets enough device memory to run its part – thinking mode off for real-time performance on budget hardware.

the language choice matters too – bisindo is the natural sign language of indonesia's deaf community, distinct from sibi, the government-imposed standard used in schools. kawanisyarat works with the language the community actually uses.

indonesia has ~2.5 million deaf citizens and only 79 schools for the deaf across the entire country. most ai accessibility tools assume you have internet. kawanisyarat assumes you don't – and still delivers a communication bridge for people who had none.