Feasibility Notes
A plain answer to whether a Gemini/Siri-style assistant can be built offline, on-device, with a local LLM.
Call recording
This is the hardest of the four goals, and the limit comes from the OS, not from AI. Since Android 9–10, third-party apps can no longer capture the other side of a normal cellular call — the accessibility-service workaround Google previously allowed is closed and banned on Play.
- Speakerphone + mic recording works with no special permissions, at lower quality.
- A phone's built-in recorder, where the OEM provides one, can be fed into the pipeline afterward.
- Desktop system-audio loopback covers VoIP and meeting apps (Meet, Zoom, Teams, WhatsApp desktop) cleanly.
Recording-consent laws vary by place. This isn't legal advice — check what applies to you before recording other people.
Wake word
Very doable. openWakeWord or Picovoice Porcupine run a small always-on detector; heavier models only load after the keyword fires.
Offline, local LLM
Doable, with size limits. Desktop hardware comfortably runs a 3B–7B model through Ollama or llama.cpp. On a phone, models are realistically capped around 1–4B, which is enough for command routing and short summaries but slower on long transcripts.
Assistant-task parity
Partial. Device actions — alarms, timers, calls, SMS, opening apps, calendar, reminders, notes, media — work fully offline through structured tool calls. Anything that needs live data (web search, weather, traffic, streaming) can't be offline by definition, so those stay optional, off by default.