Graduate Research
Multi-Agent LLM Systems & Reliability
- Built and evaluated a LangGraph Planner-Critic-Fixer pipeline with AST checks, selective reversion and MLflow tracking across five local 3B to 8B model families, ten experimental phases and three code benchmarks.
- Measured a +1.37-point average pass@1 gain across nine conditions on a 3B planner but found monitoring reduced Qwen2.5 Coder 7B from 90% to 86% pass@1, showing it can harm stronger models.
My first-author Master's research manuscript with two DePaul faculty co-authors is under conference review.
The paper, code and detailed observations are anonymized and remain private to follow the conference's double-blind review standards. The conference name is intentionally withheld.