Building a Voice-Controlled AI Agent

The ChallengeThe objective was to build a local AI agent that accepts audio, classifies intent (file creation, code writing, summarization), and executes actions on a local machine. However, as a developer working on an HP laptop with 8GB of RAM, running a full local pipeline (Whisper + Llama 3) caused significant system lag.The Solution: A Hybrid API-Local ArchitectureTo ensure a smooth user experience while meeting the assignment's safety and functional goals, I opted for a hybrid approach:STT & Intent Analysis: Offloaded to Groq Cloud using whisper-large-v3 and llama-3.1-70b for sub-second processing.Local Execution: A Python backend that manages file operations strictly within a dedicated /output folder.UI: Built with Streamlit for real-time feedback on transcription and intent.Code Deep DiveI used a "System Prompt" to force the LLM to return structured JSON, which allowed the local Python script to execute commands reliably:JSON{
"intent": "WRITE_CODE",
"filename": "hello.py",
"content": "print('Hello World')"
}

Key FeaturesSafety First: All actions are restricted to the output/ directory to avoid accidental overwrites.Graceful Degradation: The system handles API fallbacks if local resources are low.Compound Logic: (If you added this) The agent can handle complex requests like "Summarize this and save it to a file".