NewSole-author paper accepted to the NeurIPS 2026 main track → Data Scientist · Applied AI Scientist · Ph.D. Statistics

Turning ambiguous business problems into production decision systems.

Technical lead at a group holding company's AI center, previously at a telecom. LLM/RAG and agentic systems, mathematical optimization, causal inference and forecasting — owned end to end, from problem definition to deployment and validation.

See all 19 projectsAbout me
Seoul, KoreaNeurIPS 2026 · 7 patents · 9 publicationsTechnical lead, CJ AI Center
Projects · shuffled on every visit

19 projects, each with a live model

Twelve from CJ AI Center and seven from LG Uplus. Every card opens a write-up — why the project started, what it aimed for, the architecture and models behind it, and what it achieved — with the core model rebuilt on synthetic data and running in the page.

All live models →

GenAI · OCR + RAGCJ Olive Young

Ad copy legal-risk review AI

OCR with coordinates, retrieval over violation cases, an LLM judge and an independent verifier.

Precision 81% · Recall 95% · in production
Multi-agent LLMCJ Group HR

Group-wide HR recruitment assistant

Summary, evidence and interview-question agents with a self-correcting verification layer on a self-hosted model.

QA-FactEval 83.45% · review time −50%
GenAI · KG RAGSmileGut service

Dr. SmileGut personal AI advisor

Knowledge-graph retrieval, LangGraph memory and streamed answers grounded in each person's report.

Commercial launch 2025-04
Logistics · AgenticCJ Logistics

QPS allocation and injection-order engines

Knapsack-DP allocation and Tabu Search over a conveyor simulation, with an LLM floor supervisor.

+20% productivity · deployed
Customer analytics · LLMCross-affiliate

Cross-affiliate customer insight and personas

An LLM-built interest taxonomy, affinity scores, cohorts and personas across three businesses.

Drove a data-platform investment
Bio AI · EmbeddingsCJ Bioscience

Immunotherapy response and protein search

100M+ protein sequences into ~6M clusters, embeddings and a vector index searched in 1–2 s.

+15% over ML baseline · SITC 2024
Causal discoveryCJ CheilJedang

Bio-process causal optimization

Causal discovery with process-order priors and an in-silico simulator across three global plants.

≈ USD 1.86M savings · 5 patents
Logistics · MILP + DPCJ Logistics

Cart picking optimization

Order-box mapping, cart composition and routing solved as one problem with MILP and dynamic programming.

≈20% productivity in simulation
Media · CP-SATCJ CGV

Cinema screening-schedule optimization

30+ business rules in one CP-SAT model with staged relaxation, plus an occupancy forecast.

30+ rules · ≈10% MAE occupancy
Supply chain · MIQCPCJ Freshway

Network inventory placement

Storage, sourcing and pooled safety stock decided together under capacity in one model.

Capacity & transport savings quantified
Retail · ForecastingTous les Jours

Store demand forecasting and production

Year-round store × category forecasts and production recommendations for 300+ bakeries.

+5.7% revenue · licensed
Business forecastingCJ CheilJedang

Predictive management signals

Daily sales forecasts turned into a month-ahead traffic light for monthly target risk.

97.6% accuracy · ≈1 month early
Synthetic data · PrivacyLG Uplus

Synthetic data and public-sector data sharing

GAN, diffusion and copula generators under one privacy–utility QA report.

Certified-anonymous data via KISA
Causal inferenceLG Uplus

Causal drivers of mobile churn

Time-series causal discovery over ~400 indicators, then effects with PSM, IPTW and the G-formula.

Churn gap vs leader narrowed
ExperimentationLG Uplus

A/B testing platform

Sample size, randomization, frequentist and Bayesian reads and bandits for IPTV and mobile TV.

Standardized experiments
Media · ForecastingLG Uplus

IPTV movie revenue forecasting

Title totals with tree ensembles, post-release curves with the Bass diffusion model, the market with Prophet.

Acquisition & promotion on expected revenue
Customer value · SurvivalLG Uplus

Customer lifetime value

A Temporal Fusion Transformer for revenue, MTLR and Cox for retention, combined into long-term value.

Future value as the retention lens
Network analysisLG Uplus

Family and household inference

Calls, location and accounts as one graph; Parallel Louvain communities and household IDs over time.

Household-level customer view
Customer scoringLG Uplus

Engagement score and rank

RFM levels by genetic-algorithm optimal binning and least-squares service weights against ARPU and churn.

Operational fan score
More analyticsLG Uplus

Additional analytics projects

  • Preferred-team predictionfor a baseball streaming service, with BYOL and label-imbalance methods.
  • Process mining of subscription flowsto find drop-off points and fix the UI/UX.
  • Interest indices for target marketingfor LG Hausys, combined by rank aggregation.
Go deeper

Read the story behind each one

Every write-up covers why the project started, its goals, the architecture and the characteristics of the technologies used, the results — and hands you the live model.

All write-upsAll live models
19projects with write-ups and live models
7patents (lead or named inventor)
$1.86Mconfirmed bio-process cost savings
+20%logistics productivity, field-validated
9peer-reviewed papers & posters, incl. NeurIPS 2026
Experience

Where the work happened

Each project chip opens its write-up.

CJ AI Center

2023 — present
Data Scientist / Applied AI Scientist, Technical Lead

The group holding company's central AI organization, taking on cross-affiliate problems that individual companies cannot solve alone. I set technical direction, review architectures, mentor data scientists and work directly with affiliate executives and domain experts.

Ad copy legal-risk reviewR 95% HR recruitment assistant−50% review time Dr. SmileGut advisorlaunched Logistics QPS engines+20% Cross-affiliate customer insight Immunotherapy response+15% Bio-process optimization$1.86M Cart picking≈20% Cinema scheduling30+ rules Network inventory Bakery demand forecasting+5.7% Predictive management97.6%

LG Uplus

2020 — 2023
Data Scientist / Technical Lead

Customer retention, customer value, experimentation and synthetic data on telecom-scale data across mobile, IPTV and subscription services.

Synthetic data & KISA sharing Churn causal drivers~400 vars A/B testing platform IPTV movie revenue Customer lifetime value Household inference Engagement score

Also: preferred-team prediction for a baseball streaming service with BYOL and label-imbalance methods; process mining of subscription flows for UI/UX fixes; interest indices for LG Hausys target marketing combined by rank aggregation.

CCNI Research · Hanmi Pharmaceutical

2017 — 2020
Statistician / Biostatistician

Healthcare prediction and clinical-trial statistics, where survival analysis, hypothesis testing and sample-size work met real medical and pharmaceutical studies.

  • Disease exacerbation prediction — environmental and patient data for asthma, COPD and pediatric-cancer risk; SMOTE and ADASYN for rare events, deep neural networks, connected to a patient-facing app.
  • Clinical-trial statistics — sample-size calculation, analysis-method review, significance testing and statistical analysis plans.
  • Early-phase dose finding — Bayesian logistic regression, the continual reassessment method with overdose control, graphical multiple-testing procedures.
Research & IP

Publications and patents

How Invariant Hyperbolic Unfolding works: structural percentiles anchor radius, message passing learns only the angle, and an unseen graph falls onto shared radial shells
NeurIPS 2026 · Main Track · Poster · sole author

Invariant Hyperbolic Unfolding: Radial Canonicalization for Label-Free Cross-Graph Link Prediction

Hyperbolic graph encoders learn radius in each graph's own scale, so hierarchy does not transfer. Anchoring radius to structural percentiles and learning only the angle lets one frozen encoder, trained on synthetic graphs, predict links on graphs it has never seen.

+2.7 pts HR@50 over the strongest learned hyperbolic baseline · 52.5% less degradation at 50% edge removal · 4.2× gain on low-degree nodes
Read the write-upTry the live model
Invariant Hyperbolic Unfolding: radial canonicalization for label-free cross-graph link predictionNeurIPS 2026, Main Track (poster) · Lee, sole author · write-up and live model
Associations of the gut microbiome with immune checkpoint inhibitor response across cancer types and cohorts identified using random forest and language modelsSITC 2024, Poster 1265 · Yang, Oh, Lee et al. · related write-up
Statistical assessment of biosimilarity based on the relative distance between follow-on biologics for time-to-event endpointsStatistics in Biopharmaceutical Research (2020) · Lee & Kang · first-author methodology
Alertness during working hours among eight-hour rotating-shift nurses: an observational studyJournal of Nursing Scholarship 54(4), 2022
Effects of low-dose pirfenidone on survival and lung function decline in patients with IPF: results from a real-world studyPLoS ONE 16(12), 2021
Sleep, fatigue and alertness during working hours among rotating-shift nurses in KoreaJournal of Nursing Management 29(8), 2021
Readmission of high-risk discharged patients at a tertiary hospital in KoreaThe Journal for Healthcare Quality 41(4), 2019
Combined effects of diabetes and low household income on mortality: a 12-year follow-up of 505,677 Korean adultsDiabetic Medicine 35(10), 2018
Diabetes, frequency of exercise, and mortality over 12 years: analysis of the NHIS-HEALS databaseJournal of Korean Medical Science 33(8), 2018
Patents
7
Also
  • Technology-licensing agreements (demand forecasting, process optimization)
  • Internal technical documents, RL hands-on workshop, trend reports, LLM reading group
  • Teaching at Yonsei University: R & Python programming, Introduction to Statistics
About

Statistician by training, builder by habit

I hold a Ph.D. in Statistics from Yonsei University, where my dissertation developed biosimilarity assessment methodology for time-to-event endpoints. That training — survival analysis, hypothesis testing, experimental design — is still the backbone of how I evaluate every model I ship.

My core strength is redefining a business problem into an analyzable one: translating implicit domain knowledge, operational constraints and business rules into objective functions, evaluation metrics and production-ready systems. I own projects end to end, from problem definition and data pipelines through deployment and performance validation, and I prefer systems that business teams can keep using without me.

LanguagesKorean (native) · English (conversational; peer-reviewed technical writing)
ToolsPython, R, SQL, SAS, SPSS · LangGraph, vector DBs, Streamlit · MILP solvers, simulation
HonorsBrain Korea 21 Plus research excellence scholarship and fellowship · Korea Weather Big Data Contest, agency president award (2016)
2023 —
CJ AI CenterData Scientist / Applied AI Scientist, Technical Lead
2020 — 2023
LG UplusData Scientist / Technical Lead
2018 — 2020
CCNI ResearchStatistician
2017 — 2018
Hanmi PharmaceuticalBiostatistician
2015 — 2020
Yonsei UniversityPh.D. in Statistics · GPA 4.2 / 4.3
Contact

Working on a hard, unframed problem? Let's talk.

thstar.ai@gmail.com

Seoul, Korea · Always happy to talk about applied AI, optimization and causal inference problems.