Microsoft (Asia) Internet Engineering Institute
Engineering Intern · Multi-Agent AI Systems
AutoGen-Based Multi-Agent Diagnostic System
Engineered and optimized a multi-agent diagnostic system for automated error-log analysis, root-cause diagnosis, and retrieval-augmented solution generation, with integration into a Microsoft Teams bot.
- Designed an AutoGen-based multi-agent diagnostic workflow with specialized agents, tool-driven interaction, dynamic follow-up, and context backtracking for diagnosing complex distributed-job failures.
- Developed an intelligent error-log processing pipeline combining initial error localization, worker-level log routing, dynamic context expansion, noise filtering, hash-based deduplication, and structured log retention for subsequent reuse.
- Built a hybrid retrieval and solution-generation pipeline using BM25 and vector search, candidate re-embedding, MMR reranking, dynamic similarity filtering, and BREW semantic compression to retrieve diverse and relevant historical solutions.
- Optimized system efficiency and bot integration, reducing per-task context from over 20K tokens to approximately 2K and runtime from about 240 seconds to 50 seconds, while integrating the diagnostic workflow into Dri-Bot for Microsoft Teams.
Reported Impact
20K+ → ~2K tokens
~240 s → ~50 s runtime
Evaluation: 38.6% → 85.71%