05 · SEMA
Sema
Offline, on-device Kenyan Sign Language interpreter
The situation
Built for the Gemma for Good Hackathon. The challenge: Kenyan Sign Language (KSL) interpretation is a real, unmet need, but most translation tooling assumes reliable connectivity and cloud inference — both of which are real barriers for a lot of the communities that would actually use this.
Thought process
"Fully offline, runs on a phone people already own" wasn't a technical flex here — it was the actual product requirement. If it needs the cloud, it doesn't work for the people it's meant for. And it had to be two-way: sign to the app and have it interpreted for a hearing person, or have a hearing person's speech signed back — a real conversation, not a one-directional lookup tool.
Engineering decisions
Sema is a bidirectional KSL ↔ English/Swahili interpreter running fully offline, on-device, on iPhone, built on Gemma 4 (E2B) via llama.cpp — small enough to run on-device, capable enough to hold up both directions of a real conversation.
My contribution
Full build for the hackathon submission — the on-device Gemma integration and the bidirectional interpretation flow.