[GSoC 2026] Inquiry regarding Project 26: AI-Powered Diagnostic Agent for Edge Devices

[Amanjha112113]

Hello @lishanliu and @andypeng,

I am very interested in Project 26. The concept of bringing LLM-driven root cause analysis (RCA) directly to the edge—especially within the Edge Microvisor Toolkit (EMT) ecosystem—provides a solution to the "black box" problem which exists in field deployments.

After reviewing the project description, I have a few technical questions to better understand the desired scope and architecture:

1. Model Quantization and Runtime Strategy
   The agent requires native operation on EMT which results in hardware resource restrictions that affect RAM and NPU usage. Are we prioritizing OpenVINO™ for model optimization to leverage Intel hardware on these edge devices? The diagnostic agent will use a Small Language Model (SLM) which comprises Phi-3 and TinyLlama to prevent resource starvation from primary edge applications.

2. Context Window vs. Log Volatility
   Kernel logs and system journals contain enormous data volumes. When feeding these into an on-device LLM:

Will the agent implement a RAG (Retrieval-Augmented Generation) approach locally to fetch only relevant log snippets based on error signatures?

Our approach will use a sliding-window summarization technique to analyze logs after condensing them into shorter summaries.

3. EMT Isolation and Access
   The diagnostic agent needs which access level to reach every EMT partition? The agent needs access to a dedicated high-privileged management partition to conduct root cause analysis through hardware health monitoring or it will use an EMT-defined interface for communication.

4. Offline Knowledge Base Updates
   For field devices with intermittent connectivity, how do we envision the "Knowledge Base Integration"? Should we implement a lightweight local vector database which allows delta-patch updates during online periods through ChromaDB or FAISS?