
01Computer Engineering thesis2026
Lery AI
A screen-less, voice-first English tutor that runs on a Raspberry Pi.
Many learners freeze when they have to speak. A naive LLM tutor makes it worse: give it a long prompt and it answers an A1 student at B2 level. Lery treats the model as one component in a system that controls what it sees, when, and how its output is checked.
- Layered context with explicit token budgets, plus just-in-time retrieval through tool calls instead of injecting full history every turn.
- Compliance gate for beginner levels: typed violations are fed back to the tutor for one corrective retry before anything is spoken.
- Every turn scored on four CEFR pillars using schema-constrained output, with eval fixtures and a synthetic student to test the pipeline.
- OpenTelemetry traces showed two independent LLM checks running in series; running them in parallel removed a full call from the critical path.
- Role
- Design and development across the device voice loop, agent service, API and mobile app.
- Stack
- TypeScript · Fastify · Gemini · OpenTelemetry · PostgreSQL · Python · React Native











