Projects, Sungjin Choi

Soccer Player Market Value Scout

Deployed ML service that predicts and explains soccer players' market value. Streamlit app with a FastAPI prediction service on Google Cloud Run.

PythonMachine Learningscikit-learnRandom ForestRegressionFeature EngineeringModel DeploymentFastAPIREST APIStreamlitDockerGoogle CloudWeb ScrapingData VisualizationPredictive ModelingSports Analytics

Live app Code

🤖

LLM Multi-Agent Research Teams

M.S. thesis: LLM multi-agent simulations of five real research teams (~24K utterances) with a four-layer fidelity framework and a 55-metric NLP evaluation suite. Under review at EMNLP 2026.

PythonNLPLLMsMulti-Agent SystemsDeep LearningStatistical InferenceHypothesis TestingBootstrapExperimental DesignModel EvaluationData PipelineResearchReproducibility

Publication Details

🧮

Multimodal LLM Merging

Samsung Electronics AI research: layer-wise merge strategies and a scalable safetensors pipeline that lift language-benchmark scores while preserving multimodal performance.

PythonPyTorchDeep LearningLLMsComputer VisionNeural NetworksModel EvaluationResearch

Details

🔬

SLIViT Vision Transformer Reproduction

Reproduced and verified the SLIViT vision-transformer experiments from the paper and open-source code, analyzing model structure and performance across settings.

PythonPyTorchDeep LearningComputer VisionTransformersNeural NetworksModel EvaluationReproducibilityResearch

Enhancing Emotion Recognition in AI

Improved interpretability and reduced bias of an image emotion classifier using CLIP-Dissect and concept models on a ResNet-50 backbone.

PythonPyTorchDeep LearningComputer VisionCLIPTransfer LearningModel InterpretabilityNeural NetworksResearch

Live demo Code

CLIP Image Retrieval & VLM Classification

CLIP embeddings with k-NN retrieval and Qwen2-VL classification on CIFAR-10, with retrieval-quality checks against a random baseline.

PythonPyTorchDeep LearningComputer VisionCLIPscikit-learnMachine LearningModel EvaluationInformation Retrieval

Code

📊

Congressional Stock-Trading Analysis

Analysis and modeling of U.S. House members' stock trades: data collection, missingness analysis, hypothesis testing, and a Random Forest party-prediction model (99% accuracy) with a fairness permutation test.

PythonpandasData AnalysisEDAData CleaningMissing DataData CollectionHypothesis TestingPermutation TestingClassificationRandom ForestHyperparameter TuningFairnessWeb ScrapingStatistical ModelingFinance

Code

🚗

Used Car Price Prediction

Predicting used-car prices from vehicle attributes with feature engineering, exploratory analysis, and regression model selection.

Pythonscikit-learnMachine LearningRegressionFeature EngineeringEDAData AnalysisPredictive Modeling

Code

✉

Email Spam Classification

Comparing preprocessing techniques and classifiers (logistic regression, LDA/QDA, SVM, random forest) on 4,601 emails. Built in R.

RMachine LearningClassificationLogistic RegressionSVMRandom ForestStatistical ModelingFeature EngineeringData Analysis

Code

Optimizing Hospital Locations in San Diego

Geospatial analysis using traffic-crash and census data to recommend hospital locations for high-risk areas, compared against actual locations.

PythonGISGeospatial AnalysisData AnalysisData VisualizationOptimizationHealthcarepandas

Code

🎲

Gradient Boosting from Scratch

Implemented gradient boosting (XGBoost-style) from scratch with gradient- and Hessian-based optimization, matching the library's performance under the same hyperparameters.

PythonMachine LearningGradient BoostingXGBoostRegressionOptimizationscikit-learnModel Evaluation

🏢

Real-Estate Demographic Case Study

Case study joining 2023 Census demographics with LODES employment data to analyze where people live versus work, using statistical modeling and data collection from public APIs.

PythonData AnalysisStatistical ModelingData CollectionData VisualizationCensus DataEconomicsBusiness Analyticspandas

📈

Market Intelligence & Competitor Analysis

Samsung Biologics internship: market and competitor analysis to support strategy, with automated workflows that cut recurring manual effort by over 90% and insights communicated to stakeholders.

Market ResearchCompetitor AnalysisBusiness AnalyticsData AnalysisStakeholder Communication

📋

Customer Analytics Dashboards

Suntek Systems: built Tableau dashboards from customer data that cut reporting time by over 50% and delivered data-driven insights to align the product with customer needs.

TableauSQLData VisualizationDashboardsBusiness AnalyticsData AnalysisCustomer SuccessStakeholder Communication