call-assistant

High level architecture:

    Incoming Call
            ↓
    CallScreeningService (Android)
            ↓ (15 sec timer)
    If user doesn't answer
            ↓
    Auto-answer call
            ↓
    Audio Stream (Mic + Speaker)
            ↓
    Speech-to-Text (Local)
            ↓
    Local LLM (Response generation)
            ↓
    Text-to-Speech (Local)
            ↓
    Caller hears AI voice

File Structure:

            call-assistant/
            │
            ├── assistant_core.py        # Entry point (Android will call this)
            ├── conversation.py          # Call flow + state machine
            ├── llm.py                   # Local LLM (TinyLlama / Phi)
            ├── stt_whisper_stream.py    # Streaming Whisper STT (optimized)
            ├── memory.py                # Call transcript storage
            │
            ├── models/
            │   ├── tinyllama.gguf       # LLM model, this needs to be downloaded locally | because this is llm model which is really has large fiel size, so stop complaning and read this 😐 
            │   └── whisper/             # Whisper models | download this using `git clone https://huggingface.co/Systran/faster-whisper-base`
            │
            └── requirements.txt