Changyong Hyun

At MecAgent, I worked on a drawing-to-CAD pipeline that reads mechanical engineering drawings and rebuilds parts in CAD. My work centred on evaluation and experiment design: deciding whether a change actually helped.

I designed an evaluation for the drawing-reading stage using 10 reference parts built by the company's CAD expert. After finding that the benchmark missed mirrored, rotated, and imitated sheet-metal-fold errors, I built deterministic checks to detect them.

I ran controlled experiments with acceptance criteria registered before each run, and reported negative results as clearly as positive ones. When disagreements between three models exposed errors in my reference labels, I changed the process so labels came from geometry rather than being written by hand.

My projects include LoRA and DPO fine-tuning, an eval-first RAG system, and a verifier for LLM-written CAD code. I hold an MSc in Data Science from Télécom SudParis (Institut Polytechnique de Paris), and also participated in a machine-learning training program at Hanwha Aerospace, where my team's project on global defense trends, arms imports and exports, and war status placed first.