ABOUT.

Self-taught ML systems engineer. I came to engineering from digital-content production and converged into ML when I needed identity-preserving image generation for my own K-pop digital human R&D (AZIT) — locking the face of ‘Lee Seol’, a synthetic character generated from text only, with a Z-Image Turbo LoRA.

The way I work: read the field bibles (FPHam, djdante, CivitAI, HuggingFace community), turn the synthesis into a written principle document, implement the principle as code (written with Claude Code), then measure against a commercial baseline. At The Blue Studio’s BLUE LAB (staff, AI content production, March 2026) I used this method to train identity LoRAs on real actor photography and benchmark them with ArcFace against a commercial baseline (0.6050 vs 0.5090). Those documents and scripts are published with the company’s permission.

On the audio side, I run K-Pop production end-to-end (separate face at /audio). The two practices reinforce each other: audio taught me to A/B against commercial references on every decision, which is the same instinct that makes ML evaluation honest.

I am at the start of an engineering career, not in the middle of one. B.A. in Digital Content (2025), one company LoRA project (BLUE LAB) and one personal R&D project at depth, no years of web-scale backend experience.

What I do have: evidence I can scope, build, document, and measure a non-trivial ML pipeline. Nine scripts (written with Claude Code, designed and verified by me), a benchmark, a structured sweep, and a written reasoning trail for each decision. I prefer to underclaim and let the artifacts do the talking.

Best fit: mid-level individual contributor, working alongside senior engineers on AI-native product work, with room to ramp on web stack (TypeScript, React, Postgres) where projects need it.