PORTFOLIO

Selected work in Vision AI and AI.

Turning real-world visual data into structured knowledge.

01 — Lectures

Teaching Go, now produced with AI.

Two Go courses on Inflearn (33 lessons, 9 h 19 min). The new course is built from a single script into Korean and English editions, narrated in my own voice cloned with ElevenLabs.

Courses

2022.12.02

Really Easy Go for Beginners

Inflearn · 17 lessons · 4 h 21 min

Published View on Inflearn
2023.04.13

Really Easy Go: Intermediate

Inflearn · 16 lessons · 4 h 58 min

Published View on Inflearn
—

Building Fast Go Programs

Korean · English · AI production pipeline

In production
—

The next course starts right after

Next

AI Lecture Pipeline

One Markdown script becomes a finished episode. Every line is checked before any paid voice synthesis; the narration is my own voice cloned with ElevenLabs, and the slides, coding practice and motion graphics are timed to it.

Script Markdown
Script Check before paid TTS
Narration · My Voice ElevenLabs PVC · Multilingual v2
- SlidesManim · red pen
- Coding Practicecode-server · Go Playground
- Motion GraphicsBlender
Episode KO · EN

DemoSame scene, two languages

The same page from episode 1, built from one script. Both narrations are my cloned voice.

Korean

English

02 — MediaVault

An AI archive that cross-checks itself.

MediaVault ingests documents and images, extracts meaningful information with multiple AI models, and cross-checks the results to build a reliable, searchable archive.

Document Analysis

Upload
Deduplication xxHash64 → BLAKE3
Format Conversion LibreOffice → PDF
AI Analysis
- Claude
- Gemini
- Qwen3-VL
- Gemma4
Search Full-Text Search · Tag Filter

Image Analysis

Upload
Deduplication xxHash64 → BLAKE3 · dHash → pHash
Metadata EXIF
- Person DetectionYOLOv8-seg
- Age & GenderMiVOLO → InsightFace
- Face EmbeddingAdaFace
- Image EmbeddingSigLIP SO400M
Image Description
- Step3-VL
- Qwen3-VL
- Gemma4
Cross-check & Merge
- Claude
- Gemini
- Gemma4
Search Full-Text · Vector · Text-to-Image

Dedicated model
Swappable model

Models in use

Stage	Model	Role
Document analysis	Claude / Gemini / Qwen3-VL / Gemma4	Document understanding and information extraction
Image analysis (detection)	YOLOv8-seg	Person detection and segmentation
Image analysis (face)	MiVOLO / InsightFace / AdaFace	Age, gender and face embedding
Image analysis (embedding)	SigLIP SO400M	Image embedding for similarity search
Image description	Step3-VL / Qwen3-VL / Gemma4	Generate detailed image descriptions
Cross-check & merge	Claude / Gemini / Gemma4	Keep only answers agreed by two or more models

03 — FashionEngine

From product pages to searchable fashion data.

FashionEngine converts e-commerce product pages into structured fashion data, using computer vision to segment, classify, and analyze garments at scale.

Ingest

Stores the information and images provided to us, then deduplicates and indexes them for processing.

Provided Info & Images Info · Images
Deduplication Image hash
Index PostgreSQL
Processing Queue Task queue

DemoDetail Page Segmentation

A product detail page is not a stack of photos. It is one long image where a designer has blended photos, text and background decoration together. It looks like you could just cut along the white gaps, but backgrounds run on, text sits on top of photos, and designers deliberately fade edges so sections flow into each other. On top of that, JPEG compression noise scatters faint speckles across what should be clean white space, so simple rules break.

First, a fine-tuned Swin seam model draws horizontal lines across the page only where photos start and end. Text blocks and products on a plain background get no lines.

Seam: Swin seam (fine-tuned)
Training data: 6,581 images (956 real photos + 5,625 synthetic)

Sample images are AI-generated

DemoClassify

Then a fine-tuned Microsoft Florence-2 draws a box for every region and decides whether it is a person, mannequin, flat lay, fabric or text.

Only text blocks with at least one Korean word count as text; English-only lettering is treated as decoration. Photos are boxed along clear edges, or around the smallest area of the garment when there is no edge.

We hand-labeled 2,625 labels in Label Studio to build the training data.

person
text
person
person
text
text
text
flat_lay
text
flat_lay
flat_lay

Judge: Florence-2 (fine-tuned) · 5 classes
Training data: 2,625 labels, hand-made in Label Studio
OCR: Qwen3-VL 8B

Sample images are AI-generated

DemoGarment Masking

SAM3 takes a keyword from the product name (e.g. knit → "knit sweater") and segments only the target product, pixel by pixel. It follows the garment's outline instead of drawing a box; everything else is made transparent and the result is cropped. Only this cut-out goes on to embedding.

Model: SAM3 (text prompt · knit sweater)
Next: Only the cut-out goes to embedding

Sample images are AI-generated

Embed

Turns each garment into vectors to find similar items.

Model: DINOv3 ViT-L (1024-d) · Marqo-FashionSigLIP (768-d)
Search: pgvector (HNSW, cosine)

Analyze

Reads the garment's attributes from representative shots.

Inter-VLM Agreement Verification

Three VLMs judge each property separately, and an LLM settles the answer by majority vote. When there is no majority, a person reviews it (QBC, Query-by-Committee).

Representative Shots Representative shots
- VLM 1Property vote
- VLM 2Property vote
- VLM 3Property vote
Majority Vote LLM · majority vote
- Confirmed2 of 3 agree
- Human ReviewNo majority · QBC

Judges: VLM × 3
Decision: LLM majority vote · no majority → human review (QBC)
Attributes: Color · Pattern · Detail · Body type and dozens more

04 — Kinevo

Learning posture and motion in 3D.

Coming soon.

A mannequin swinging a golf swing without a club inside a 3D motion viewer