AZIT

JUNGCHAN LEE

ML SYSTEMS ENGINEER

Self-taught ML engineer. Build production Python pipelines for LoRA fine-tuning, parameter optimization, and identity evaluation. Sole author of the AZIT LoRA pipeline (independent R&D for K-pop digital human) — nine Python scripts, training on RTX 5090, evaluated against a commercial baseline (Higgsfield Soul 2.0). Most recently AI Lab Lead at The Blue Studio, applying the same techniques on real actor photography.

KEY METRIC  ·  KEY RESULT
0.6050 ArcFace  ·  vs 0.5090 baseline
AZIT LoRA on FLUX.2 Klein 9B  · +18.9% OVER BASELINE
MOST RECENT ROLE

AI Lab Lead

The Blue Studio — sole engineer

PERIOD
March 2026 — April 2026
SCOPE
LoRA training · inference · pipeline build
STACK
Python · ComfyUI · AI-Toolkit · ArcFace
STATUS
Open to roles
WORKING ENVIRONMENTS

Tools that ship every dependency

We capture how the AZIT pipeline runs — the dataset preparation step in AI-Toolkit, and the inference graph in ComfyUI. Full step-by-step trace in VISUAL WORKFLOW ↗.

ComfyUI inference workflow — LoRA load, dual CLIP encode, KSampler, FaceDetailer (Impact Pack)
ComfyUI inference workflow — LoRA load, dual CLIP encode, KSampler, FaceDetailer
HONEST FRAME
B.A. in Digital Content (DCDC, 2025), self-taught ML engineering through one production project at depth. I do not have years of ML-board-backed experience; I do have evidence I can scope, build, ship, and measure a non-trivial ML pipeline solo — nine scripts, a benchmark, a structured sweep, a visual workflow trace — doing the talking instead of credentials.