Docs / Product
Technical direction
Proposed components. None are integrated yet.
The plan is to build and validate the assistant in an Android emulator, then investigate integration with a compatible Android system image on the target phone.
| Area | Proposed approach |
|---|---|
| Interface | Kotlin and Jetpack Compose for the chat launcher and Peter settings |
| Commands | Direct parser for common actions; validated model tool calls for complex requests |
| Local inference | llama.cpp with one small quantized model (Qwen2.5-1.5B Instruct is the first candidate to evaluate) |
| Voice | openWakeWord, whisper.cpp and Piper-compatible speech via sherpa-onnx are candidates |
| System | Android intents and platform APIs; privileged integration evaluated separately |
| Resources | Lightweight listener plus a separate on-demand AI process |
These are proposed choices. Device support, model and voice licenses, memory use, latency and battery impact must be checked before components are selected for a release.