Minh Duy User

I've built CogniEDA, a validity-preserving research-state infrastructure for long-running analytical investigations, with typed state, SQLite persistence, provenance controls, and durable execution recovery. I also developed EXACT, a neuro-symbolic STEM question-answering pipeline that reached the Top-10 finals among 180 IEEE IJCNN 2026 challenge teams.

I improved first-order logic exact-match accuracy from 38.5% to 75.3% through Qwen3-8B PEFT/LoRA fine-tuning, token-level loss weighting, and synthetic multi-hop data. In Datathon 2026, I ranked 49th of 527 teams by building a LightGBM-based revenue and COGS forecasting pipeline that reduced the composite score by approximately 36%.