Homai

A Bashkir-speaking smart speaker.
A real-time voice agent: on-device wake word → ASR/TTS workers → an LLM agent with actions and RAG over user data.

I’m the founder of Homai.

What matters here

• Real-time: voice UX needs predictable latency and stable behavior.
• Agentic actions: the assistant can execute tasks via tools/workflows.
• RAG: answers grounded in user-specific data.
• Production engineering: separate workers/services to control latency and scaling.

Architecture (high-level)

• On-device (C++): wake word and core logic.
• Backend (Go): orchestration, state, queues/workers.
• Workers: ASR / TTS / LLM as independent services.
• LLM agent: answers + actions (routines, automation).
• RAG: retrieval over user data.

Links

• Project site: homai.ru
• Homai Tech: homai.tech
• AI for Good: aiforgood.itu.int

Manufacturing & quality control

We manufacture devices in-house. Our SMD assembly machine runs software we wrote; the pipeline uses ML models and a vision LLM to improve assembly quality and reduce errors.

Media