Woojin (Jay) Park · Selected research, projects & systems

AI research and applied science, from original methods to real-world systems.

I am a doctoral AI researcher and product builder with more than a decade of experience across academia and industry. This portfolio brings together original research, partner projects, and deployed systems spanning recommender systems, language AI, healthcare, manufacturing and SCM, and public-interest technology.

01 / Core research

Measurement, ranking, and reliable language AI.

Four connected studies: audit platform traces, preserve evidence in ranking, route uncertain language predictions to selective reasoning, and test history-wide agreement in frozen news rankers.

PEC Audit visual showing preference, exposure, interface surface, and consumption as distinct traces
R1Algorithmic auditing · recommender systems
Korea UniversityDoctoral researchinsighteaPlatform data

When Exposure Is Not Attention: Auditing the Preference–Exposure–Consumption Gap in Personalized News Recommenders

A measurement framework for separating what a recommender stored from what a user may have noticed, chosen, or read.

Trace model

Keeps preference, logged exposure, interface surface, and consumption separate—not interchangeable proxies.

Audit protocol

Tests same-day traces and profile–click alignment using cohort rules, entropy, HHI, Jensen–Shannon divergence, and matched-count controls.

Inference boundary

Identifies where retained lists cannot establish attention, reading depth, or causal recommendation effects.

  • Six months of deployed mobile-news logs
  • Sole-authored research and public artifact
  • Interactive measurement tutorial
Read paperGitHubInteractive tutorial
PEC-Fuse source-preserving evidence streams and candidate ranking
R2News ranking · evidence fusion
Korea UniversityDoctoral researchinsighteaPlatform dataMicrosoftMIND dataset

PEC-Fuse: Source-Preserving Evidence Fusion for News Ranking with Partial Recommendation Traces

A candidate-scoring architecture that preserves where evidence came from when recommendation histories are partial.

Source-preserving representation

Keeps preference, exposure, and consumption separate so missing or conflicting traces stay visible.

Late evidence fusion

Combines signals at scoring time instead of compressing the user journey into one history vector.

Candidate-aware ranking

Scores evidence for each candidate, exposing provenance and behavior under partial histories.

  • Executable ranking artifact
  • Interactive method tutorial
  • IEEE Access manuscript in preparation
Live tutorial
Calibrated selective LLM pipeline with separate selection, correction, and full-set audit layers
R3IEEE Access manuscript · Healthcare AI · selective LLMs
Korea UniversityDoctoral researchRedditPlatform data

Confidence-Guided Selective LLM Re-Evaluation for Emotion Classification in Social Media

A calibrated cascade that reserves structured LLM reasoning for uncertain social-media emotion predictions while keeping selection, correction, and cost separately auditable.

Risk-bounded routing

Temperature-scales DistilBERT and fixes a maximum-coverage threshold on validation data under a one-sided selective-risk bound.

Structured re-evaluation

Compares Direct, Chain-of-Thought, and task-adapted SELF-DISCOVER with fixed checks for subject, time frame, competing evidence, and alternatives.

Audited evaluation

Tracks coverage, captured errors, corrections, and new errors across Reddit, mixed-emotion, and same-source tests.

  • Fixed 9,000-post evaluation
  • 1.23% routed to the LLM
  • 97.58% → 97.77% · Holm-adjusted p = .0412

Scope: three-class, non-clinical proxy emotion classification; the study does not make individual diagnosis or treatment claims.

Project pageInteractive tutorialGitHub
The same click history supplies candidate-weighted context and uniform history-wide agreement; matched final heads refine a frozen score within a bounded correction.
R4Planned ACM SIGIR submission · January
Korea UniversityDoctoral researchMicrosoftMIND dataset

Are Candidate-Weighted History Summaries Enough? Bounded History-Agreement Refinement for Frozen News Rankers

Tests whether uniform agreement across a reader’s recent clicks adds ranking information beyond candidate-weighted context, using a bounded correction to a frozen scoring pipeline.

Two views of one history

Contrasts candidate-weighted context with mean candidate–click cosine similarity across nine semantic fields, using the same last 30 clicks.

Matched refinement

Freezes the backbone and semantic adapter, then fits identical final-head architectures from paired initial states. Only the added history representation changes.

Bounded score changes

Caps each final correction at ±0.05 on the supplied score scale. Pairwise margins above 0.10 preserve order; the bound alone does not guarantee relevance gains.

  • MIND-small · 73,152 evaluation impressions
  • NAML · NRMS · CAUM · three matched seeds
  • ΔnDCG@10 +0.000978 / +0.001422 / +0.002853 vs. same-head control

Evidence: positive user-cluster 95% intervals across all nine metric–backbone comparisons, conditional on the fitted seeds. Results cover one offline benchmark.

Publication record

02 / Applied partnerships

Methods shaped by real partners and constraints.

Each project connects a real partner problem to its evidence, method, and decision context; expandable notes clarify role and inference boundaries.

Supplier invoices pass through AWS Textract and Google Document AI, then field matching and extraction coverage checks produce structured records
F1Document AI · cloud automation
Carnegie Mellon UniversityCapstone · Systems administration & ML developmentEatonIndustry partner

Intelligent Invoice Processing with Eaton

CMU industry capstone converting varied supplier invoices into structured records for Eaton’s financial shared-services workflows.

  • ExtractionAWS Textract and Google Document AI for invoice fields and line-item tables.
  • IntegrationAWS pipeline, field-alias matching, and standardized CSV outputs.
  • EvaluationCross-provider mismatch detection and extraction-coverage scoring.
My role & project evidence

Systems Administrator / ML Development on a five-person team advised by Prof. Neelam Dwivedi, working with Eaton IT and Global Financial Shared Services. The team presentation reports 388 tested invoices, 91% extraction coverage for common fields, and 69% for table fields. These are project-level results; extraction coverage is distinct from OCR accuracy. Human review for low-coverage invoices is described as a next step.

Open presentation & technical report
Applied mental-health AI research path from behavioral data and BERT to stakeholder review and selective LLM reasoning
F2Mental-health AI · applied language modeling
Carnegie Mellon UniversityResearchNeolthIndustry partner

Mental-Health Risk Modeling at Neolth

Language modeling for youth mental-health risk signals and professional review.

  • DataReddit research corpus and longitudinal behavioral signals.
  • ModelingNMF topic discovery and BERT classification; proposed active-learning refinement.
  • TranslationRisk questions and outputs informed by school stakeholders.
Research boundary

These models support risk-oriented research and review; they are not clinical diagnoses.

Open project proposalCurrent research
Longitudinal patient journeys, engagement cohorts, and potential outcomes analysis
F3Healthcare · longitudinal and causal analysis
Carnegie Mellon UniversityApplied projectClínicas del AzúcarHealthcare partnerDeloitteProject collaboration

Improving Patient Engagement at Clínicas del Azúcar

CMU and Deloitte project on diabetes-care engagement and personalized intervention design.

  • MeasurementMonthly engagement trajectories from patient and activity records.
  • ModelingK-means/GMM segmentation and random-forest engagement prediction.
  • Decision designObservational causal analysis to inform a proposed nutrition and activity intervention.
Evidence boundary

The causal analysis estimates an intervention scenario from observational data; it is not evidence from a randomized clinical trial.

Open work sample
Terrain, remote sensing observations, and policy pathway for North Huntingdon landslide mitigation
F4Public-interest technology · municipal data strategy
Carnegie Mellon UniversityPolicy Innovation LabNorth Huntingdon TownshipMunicipal partner

Municipal Landslide Mitigation and Prediction

CMU Policy Innovation Lab project on landslide mitigation for North Huntingdon Township.

  • Evidence reviewMunicipal records, GIS, risk models, and LiDAR/SAR options.
  • Stakeholder researchExpert interviews, literature, and local ordinance comparisons.
  • DeliverableRecommendations for open data, slope-aware policy, and agency coordination.
Role

Data Analyst on a four-person team connecting evidence collection, infrastructure risk, and feasible policy intervention.

Open reportPublic case

03 / Products & systems

Research translated into working decisions.

Deployed recommendation, patented interfaces, and industry ML systems with a concrete operational purpose.

Conceptual workflow: behavior, search intent, and restaurant embeddings feed hybrid ranking for personalized discovery
S1Production recommendation
CatchTableStaff Data Scientist · Recommendation & ranking

CatchTable: Recommendation, Ranking, and Personalized Search

Led user profiling and hybrid recommendation, connecting behavioral signals and search intent with restaurant embeddings.

  • GPU-accelerated clustering · LLM representations
  • Vertex AI · Kubeflow training and serving
  • ~30% higher reservation conversion: team result reported by Google Cloud
Google Cloud case
Illustrative AI Equalizer: topic, publisher, diversity, and recency controls adjust a personalized news feed
S2Patented recommender interface
insighteaCo-founder & Chief AI Scientist · Personalized news

AI Equalizer: User-Controllable News Ranking

Patented controls for topic, publisher, diversity, and recency in real-time news ranking.

  • Preference-to-ranking feedback
  • Multi-objective feed control
  • Korean patent KR102965849B1
Open patent
Conceptual workflow: profile text and activity signals become topic and vector representations for career partner ranking
S3Hybrid recommendation · patented methods
CoffeechatCTO · Recommendation systems

Career Partner Matching with Behavioral Signals and Topic Models

Combined activity-log interests and profile text with topic models and vector-similarity ranking.

  • Activity-log interest extraction
  • LDA, NMF, LSI, and BERTopic
  • TF-IDF · dense vectors · PCA
  • Two granted Korean patents
Patent 1Patent 2
Conceptual workflow: protocol and candidate data feed trial-interest ranking for outreach team review
S4Healthcare decision intelligence
IQVIASenior AI Scientist · Clinical-trial intelligence

Clinical-Trial Recruitment Modeling and ML Pipelines at IQVIA

Led delivery across a ten-person international team, developing trial-interest scores and candidate rankings to help outreach teams prioritize recruitment.

  • Protocol-informed features · imbalanced classification
  • Spark model comparison · cross-validation
  • Automated scoring pipelines: AWS · Cloudera ML · Airflow
Illustrative arrival forecasts translated into required dock-door hours for capacity planning
S5Forecasting · operations research
CoupangSenior Data Scientist · Forecasting & optimization

Inbound Inventory Forecasting and Dock-Capacity Planning at Coupang

Forecast actual SKU-level arrivals from purchase orders and supplier delivery patterns; translated truck-arrival forecasts into dock-door hours for operations dashboards.

  • SARIMAX · Prophet–LightGBM ensembles
  • Bayesian estimates for purchase-order status
  • AWS slot-requirement pipeline: 15% lower forecast error

04 / Technical projects

Original reports, grouped by method.

Original reports and posters spanning language analysis, prediction, computer vision, manufacturing data systems, and automotive statistics.

01

Language & predictive modeling

2020NLP Analysis of Consumer Posts: Samsung Galaxy S8, iPhone 8, and iPhone XBuilt a 500K+ post analysis pipeline using NMF topics, sentiment, release-window comparisons, and time-series validation to distinguish product-specific discussion patterns.Open PDF report 2019Washington D.C. Capital Bike Sharing Usage PredictionBenchmarked decision tree, random forest, and GBDT models on 17,397 observations; translated the best model's feature importance into commute- and weekend-capacity actions.Open PDF report 2019Airbnb Host Rating Optimization ModelCompared linear and tree-based regressors on 49,057 listings, tuned ensemble models, and surfaced host-controlled features associated with review scores.Open PDF poster 2018Airline Customer Satisfaction ModelingCombined explanatory regression, association analysis, and cross-validated SVM classification across 129,889 surveys to isolate operational satisfaction drivers.Open PDF report
02

Data systems & ML foundations

2018Car Manufacturing Process Monitoring SystemDesigned a normalized MES–HR relational system and SQL decision queries for labor-hour variance, production shortfalls, and factory-to-component traceability.Open PDF report 2020Pedestrian Detection with Mask R-CNNImplemented and evaluated an instance-segmentation pipeline combining region proposals, bounding-box classification, confidence and IoU checks, and per-instance masks.Open PDF report 2017Research and Findings on the Relationship Between Horsepower and Engine SizeAnalyzed 92 vehicles using hypothesis tests, chi-square analysis, and linear regression; assessed confidence intervals, distributional assumptions, and residual diagnostics.Open PDF report
See the full professional timeline

Selected chronology

A continuous research-to-production arc.

Technical foundationsData systems, predictive modeling, computer vision, and public-interest technology
Healthcare AIMental-health language modeling, patient engagement, and clinical-trial intelligence
Deployed decision systemsPatented matching, supply-chain forecasting, recommendation, ranking, and search
Doctoral researchMeasurement audits, evidence fusion, selective LLMs, and bounded ranker refinement

Contact

Research and technical collaboration

I welcome collaborations where careful evaluation and practical implementation both matter, particularly in recommendation and language systems, healthcare, and public-interest technology.

Contact Woojin