"Always-On" Deepseek AI Assistant

A pattern for an always on AI Assistant powered by Deepseek-V3, RealtimeSTT, and Typer for engineering

Checkout the demo where we walk through using this always-on-ai-assistant.

Setup

• cp .env.sample .env
• Update with your keys DEEPSEEK_API_KEY and ELEVEN_API_KEY
• uv sync
• (optional) install python 3.11 (uv python install 3.11)

Commands

Base Assistant Chat Interface

See main_base_assistant.py for more details. Start a conversational chat session with the base assistant:

uv run python main_base_assistant.py chat

Typer Assistant Conversational Commands

See main_typer_assistant.py, modules/typer_agent.py, and commands/template.py for more details.

• --typer-file: file containing typer commands
• --scratchpad: active memory for you and your assistant

• --mode: determines what the assistant does with the command: ('default', 'execute', 'execute-no-scratch').

• Awaken the assistant

uv run python main_typer_assistant.py awaken --typer-file commands/template.py --scratchpad scratchpad.md --mode execute

1. Speak to the assistant Try this: "Hello! Ada, ping the server wait for a response" (be sure to pronounce 'ada' clearly)

2. See the command in the scratchpad Open scratchpad.md to see the command that was generated.

Assistant Architecture

See assistant_config.yml for more details.

Typer Assistant

See assistant_config.yml for more details. - 🧠 Brain: Deepseek V3 - 📝 Job (Prompt(s)): prompts/typer-commands.xml - 💻 Active Memory (Dynamic Variables): scratchpad.txt - 👂 Ears (STT): RealtimeSTT - 🎤 Mouth (TTS): ElevenLabs

Base Assistant

See assistant_config.yml for more details. - 🧠 Brain: ollama:phi4 - 📝 Job (Prompt(s)): None - 💻 Active Memory (Dynamic Variables): none - 👂 Ears (STT): RealtimeSTT - 🎤 Mouth (TTS): local

Resources

• LOCAL SPEECH TO TEXT: https://github.com/KoljaB/RealtimeSTT
• faster whisper (support for RealtimeSTT) https://github.com/SYSTRAN/faster-whisper
• whisper https://github.com/openai/whisper
• examples https://github.com/KoljaB/RealtimeSTT/blob/master/tests/realtimestt_speechendpoint_binary_classified.py
• elevenlabs voice models: https://elevenlabs.io/docs/developer-guides/models#older-models