Buzo Architecture
Hybrid edge + cloud system for voice, vision, memory and physical-world action.
System Design
A concrete view of the systems behind Buzo: device onboarding, edge inference, agent memory, cloud orchestration and sensor-aware workflows.
Hybrid edge + cloud system for voice, vision, memory and physical-world action.
Local short-term state plus semantic long-term memory for agent continuity.
Tool-aware planning loop that decides when to answer, remember, alert or call an integration.
Low-latency local models for wake, transcription, detection and first-pass interpretation, evolved from the earlier pocket_ai R&D pipeline.
Escalates expensive reasoning, knowledge retrieval and integrations to cloud services.
Phone provisions Wi-Fi, identity and device settings over BLE before the device joins cloud flows.
A modular stack for camera, microphone and future environmental sensors.
Native Android agent loop proving mobile assistant primitives before they move into Buzo hardware.
Software stack exploration for Qwen3-8B on hardwired analog compute-in-memory silicon.
Voice request
Wake word -> Whisper Tiny -> intent routing -> local memory -> cloud LLM when needed -> spoken response
Vision alert
Camera frame -> YOLO11 detection -> spatial reasoning -> edge rule -> Telegram alert -> memory event
BLE setup
Mobile app -> BLE discovery -> credential transfer -> device registration -> cloud sync
Memory write
Conversation event -> summarizer -> embedding -> local index -> cloud backup policy
Android hands
User request -> LLM tool call -> accessibility read/tap/type/swipe -> result returned to agent loop
Hardwired Qwen path
FP16 Qwen3 -> 3-bit ROM quantization -> golden simulator -> LoRA fusion -> OpenAI-compatible serving