Projects

Write-ups, with the model running in the middle

Each post explains what the problem was, how the model decides, and what it changed — and embeds a re-implementation of the core on synthetic data so you can run it. Twelve from CJ AI Center, seven from LG Uplus.

Featured · GenAI · Regulated domain · In production

Reviewing ad copy for legal risk — with the reason and the location

OCR with coordinates, retrieval over violation cases, an LLM judgment and an independent verifier. Precision 81.14% and recall 94.96% on 499 ad images; patent as lead inventor.

October 2026 · 8 min · live model in the browser
GenAI · Multi-agent · Self-hosted

A recruitment assistant that checks its own work

Parallel summary, evidence and question agents with a self-correcting verification layer.

QA-FactEval 83.45% · review time −50%
GenAI · RAG · LangGraph

A personal health advisor built on validated knowledge

Report-grounded answers, hybrid retrieval and workflow control for a commercial service.

launched commercially 2025-04
Logistics · Simulation · Agentic

Two engines for a ring-conveyor picking system

SKU allocation with affinity and knapsack DP; injection order by simulation and Tabu Search.

+20% productivity · deployed
Manufacturing · Causal inference

Finding the conditions that actually move yield

Causal discovery with process order as prior knowledge, and an in-silico simulator.

≈ USD 1.86M savings · 5 patents
Logistics · Optimization

Cart picking as one optimization problem

Assignment, routing and bottleneck waiting solved together under a zone → aisle → travel objective.

≈20% productivity in simulation
Media · Integer programming

Thirty rules and one day of screenings

Business rules as constraints of one model, plus occupancy prediction.

30+ rules · ≈10% MAE
Retail logistics · Optimization

Placing stock across a food-distribution network

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

feasibility study
Customer analytics · LLM

A common language for customers across businesses

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

drove a data-platform investment
Bio AI · Protein embeddings

Microbiome signals of immunotherapy response, and a search engine for proteins

100M+ sequences into ~6M clusters, a vector store searched in 1–2 s, and a response model with biomarker candidates.

+15% over baseline · SITC 2024
Retail · Demand forecasting

From a Christmas-cake forecast to every store, every day

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

+5.7% revenue · licensed
Business forecasting · Signals

Will we hit this month's target? A month before the books close

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

97.6% accuracy · ≈1 month ahead
Causal inference · Time series · LG Uplus

Finding what leads churn, not what moves with it

Time-series causal discovery over ~400 indicators, then effect estimation with matching, weighting and the G-formula.

≈400 variables · churn gap vs leader narrowed
Synthetic data · Privacy · LG Uplus

Data that can leave the building

GAN, diffusion and copula generators under one privacy–utility QA report; certified-anonymous data provided through KISA.

certified anonymous · 3 datasets
Experimentation · Statistics · LG Uplus

An experimentation platform for IPTV and mobile TV

Planned samples, randomization, intervals and p-values, Bayesian probability and bandits — for product teams.

controlled experiments, standardized
Media analytics · Forecasting · LG Uplus

Forecasting a film's VOD sales, and the market it sells into

Title totals with tree ensembles, trajectories with the Bass diffusion model, the market with Prophet.

three forecasts · acquisition & promotion
Network analysis · LG Uplus

Who lives with whom: household inference from network signals

Calls, night-time location and accounts as one graph; Parallel Louvain, strong-tie households, IDs that last.

household view · bundles & family campaigns
Customer value · Survival · LG Uplus

Worth keeping: customer lifetime value from survival and revenue

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

future value replaces ARPU as the lens
Customer analytics · Scoring · LG Uplus

Who are the true fans? An engagement score and rank

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

operational score & rank