Gyunpyo LeeArgumentsNextRésuméGitHub
  1. CCS Detroit transportation design
  2. Changan Turin exterior design
  3. KAIST Daejeon design + deep learning
  4. LG Seoul data science and AI
papers
12
one first author · CVPRW 2022
grounded answers
77/80
10 questions × 8 runs
years, drawing to data
15.0
design 8 yrs · data 6 yrs
side profiles
1,480
GP22 · 37 brands · 10 segments
seoul
Seoul time

Gyunpyo Lee

Senior Researcher · LG Electronics

이건표 · Seoul · trained as a car designer

I design AI products and measure whether they work.

GitHub · Résumé · LinkedIn
gyunpyolee@gmail.com
gyunpyolee.com · Seoul

The bottleneck in applied ML is rarely the model.

It is the distance between the model and the customer who has to act on it. I build production LLM and prediction systems and then embed with the internal teams that depend on them. Automotive exterior design in Turin, HCI research at KAIST, production ML in consumer electronics — the through-line is that I was the customer I now build for.

2022–now Seoul

Senior Researcher

LG Electronics, Business eXperience Informatics

  • Built a multi-agent analysis system over survey data and its routing layer. When FAISS retrieval proved too slow for the full dataset, I replaced it with agents defined from metadata keywords and sent every numeric question to deterministic code. It now sits behind internal reports and leadership presentations.
  • Built, in a one-month sprint with HR, an MVP of a synthetic persona system for pre-flight focus-group interviews. The direction grew into a larger employee-research project, now carried forward by an external firm.
  • Designed and built the core of a multi-pillar customer data intelligence platform within a team: 26,000 respondents across Korea, the United States and Vietnam, 3,500+ variables, a JSON-based store designed for analyst self-service.
  • Purchase prediction at 83% accuracy, life-event forecasting at 87%. Both consumed directly by marketing and CX teams, not by a dashboard.

2022–2024 alongside

AI Consultant, Code V

  • YOLOv5 defect detection across three product categories: 98% precision, mAP@50 of 87%. Custom augmentation for the long-tail defect classes that made the baseline unusable in production.

Chief Product Officer, Dream2Real

  • Synthetic data generation for data-scarce domains, with digital twin environments in Unreal Engine and Blender where real capture was not viable.

2020–2022 Daejeon

Design Researcher, MS Industrial Design

KAIST

  • GP22, first author, CVPR Workshops 2022: 1,480 side-profile images across 37 brands and 10 segments, with a YOLOv5 design-feature detector. Dataset public on Zenodo.
  • Kia Design Center, PBV active lighting: interior lighting scenarios and VR-based UX testing with Kia's own designers as participants.
  • LS Automotive, affective lighting: EEG-based validation of emotional lighting scenarios. Published in IEEE Access.

2016–2019 Turin

Exterior Designer

Changan European Design Center

  • Key exterior designer on Changan's first platform-based vehicle (UNI-V); owned surface refinement and the handoff between CAD and clay modelling teams.
  • Facelifts for production sedans (C301, C201). Eado XT (C211) sketches on the cover of Auto & Design, 2018.
  • Exterior contributions to the Jiliu and Yuyue concepts, Shanghai Auto Show 2017.

2011–2015 Detroit

BFA, Transportation Design

College for Creative Studies

Selected publications

  • GP22: A Car Styling Dataset for Automotive Designers. G. Lee, T. Kim, H.J. Suk. CVPR Workshops, 2022.
  • User responses to dynamic light in automobiles with EEG and self-assessments. T. Kim, G. Lee, M. Park, H.M. Lee, J.W. Park, H.J. Suk. IEEE Access 10, 2022.
  • Sketching in-vehicle ambient lighting in virtual reality with the Wizard-of-Oz method. T. Kim, A. Shunayeva, G. Lee, H.J. Suk. Digital Creativity 33 (1), 2022.
  • Full list on Google Scholar

Honors

Best Paper Award, HCI Korea 2022. Finalist, Ferrari Top Design School Challenge.

Works in

LLM systems RAG pipelines, multi-agent routing, deterministic tool execution, evaluation harnesses. ML Python, PyTorch, scikit-learn, LightGBM. Vision YOLOv5, synthetic data in Unreal Engine and Blender. Design Autodesk Alias, Blender, Keyshot.

Languages

Korean native, English bilingual, Italian conversational.

Full résumé as PDF, on request LinkedIn gyunpyolee@gmail.com

I build data products the way a single stroke of light sets a volume in car design: nothing that doesn't carry the argument stays on the page, and no figure that can't be checked goes in the answer.

Arguments

Each one is a claim, a build that tests it, and the measurement. The failures are part of the record.

2026 · LIVE

Who did the math?

When an LLM does the analysis, it invents the numbers. When it only narrates numbers you computed, it doesn't. Ten questions on a synthetic 10,000-household vehicle panel, three ways of answering, every number checked — eight runs across five labs. The naive-grounded model read individual records perfectly and got every aggregate wrong. The same model run twice moved more than two different labs did.

Interactive pageRepo — every prompt and answer committedreadback — the check as a tool
Ungrounded
50%
Naive-grounded
61%
Pipeline-grounded
99.7%
Share of stated numbers that match the data. Ten questions, eight runs, six model configurations, 1,719 numbers checked.

2022 · 2026 · IN PROGRESS

Proportion, by segment

GP22: 1,480 car side profiles across 37 brands and 10 segments, with a detector for the six design features that define a profile — body, bodyside, cabin, daylight opening, wheel, tire — at 0.995 mAP. First author, CVPR Workshops 2022. Now an interactive proportion tool: one side-profile line, the six GP22 boxes, every length in tire diameters, fourteen real cars as presets you can morph between.

Paper, arXivDataset, ZenodoOpen the tool
BMW 3 Series · D segment · traced, not drawn L = 7.13 T · H = 2.23 T

Check a sentence

The check from the first argument, in your browser. Edit the sentence; every number is looked up in the statistics on the right.

Statistics the sentence is checked against

  

  
One sentence against fixed statistics here. With your own answer and your own source, and a tolerance gauge for every number: readback.






  

    
Planned
Timeline-first synthetic panels
The generator, generalized: customers as lives rather than rows, so that causal questions have real answers to test against.

  






  
Elsewhere

  

    GitHub · glee2dev
    Résumé
    LinkedIn
    gyunpyolee@gmail.com