Current Project
Samsung Health × Insurance AI
LangGraph
Claude API
FastAPI
Next.js
Transforming life insurance from a static contract into an active health partnership. Built under Samsung Lifenology Lab's entrepreneurial program with direct funding and mentorship. The system continuously ingests Samsung Health wearable signals and orchestrates a multi-agent AI pipeline to score longitudinal risk, surface personalized health interventions, and dynamically align coverage with real health trajectories — incentivizing better habits through premium reductions.
- Multi-agent AI pipeline using LangGraph to orchestrate health risk assessment, intervention recommendation, and coverage optimization workflows.
- Real-time ingestion of Samsung Health wearable data (activity, sleep, heart rate) with anomaly detection and trend analysis.
- Dynamic premium adjustment engine that aligns insurance terms with longitudinal health behaviors, creating economic incentives for wellness.
- Full-stack deployment: FastAPI backend for ML/AI orchestration and Next.js frontend for member dashboard and health insights.
Active Entrepreneurial Initiative
Samsung Health × Insurance
In Progress
Program
Samsung Lifenology Lab
Student-Led Funded Initiative
Student-Led Funded Initiative
Focus Areas
AI/ML Architecture, Wearable Integration, Multi-Agent Systems
Completed Projects
Sparse Attention CUDA Kernel
CUDA
C++
PyTorch
- Wrote a sparse attention kernel in CUDA with Longformer style masking (local window + global stride tokens); tiles with no active pairs are skipped entirely, bringing complexity from $O(n^2)$ down to $O(n \cdot w)$.
- Initial version stored a full score array per thread scaled with sequence length which spilled to global memory at longer sequences; rewrote with Flash Attention style online softmax, dropping per-thread memory to $O(1)$ and achieving $1.4$--$1.8\times$ speedup.
- Wrapped kernels as a PyTorch extension via pybind11 and benchmarked against PyTorch dense attention on a T4 GPU across sequence lengths $64$ to $512$.
CUDA · GPU Kernels
Sparse Attention Kernel
sparse pattern
local
global
skip
1.8×
v2 speedup
over v1
over v1
O(n·w)
sparse
complexity
complexity
O(1)
per-thread
memory
memory
v1 → v2 key changes
score storage
full array → streaming
softmax
materialized → online
register spill
yes → eliminated
Volatility Inference with SDEs & Data Assimilation
SDE
Kalman Filter
Particle Filter
Estimated cryptocurrency rolling volatility using a mean-reverting stochastic differential equation with online Bayesian filtering. Results were benchmarked against GARCH(1,1) and GARCH(2,2) models under a strict out-of-sample evaluation setup.
SDEs · Bayesian Filtering · Crypto
Rolling Volatility Inference
with Data Assimilation
with Data Assimilation
Method Comparison
Methodology Summary
1) Model latent volatility as a mean-reverting stochastic state.
2) Estimate the state online using Kalman and Particle Filter updates.
3) Benchmark final estimates against standard GARCH baselines.
2) Estimate the state online using Kalman and Particle Filter updates.
3) Benchmark final estimates against standard GARCH baselines.
Yut AI - Korean Traditional Board Game AI
Python
Game Theory
Bayesian Optimization
Developed strategy bots for Yut, a traditional Korean board game competition. Tested minimax tree search and heuristic-based strategies. Final approach used heuristic evaluation with Bayesian optimization (Gaussian Process + UCB) for weight tuning. Consistently outperforms baseline strategy with 54-56% win rate.
Stochastic Adversarial Game · MDP
Yut AI
stochastic adversarial game · Markov Decision Process variant
win rate vs baseline
baseline (50%)
56.5%
best win
rate
rate
8
heuristic
weights tuned
weights tuned
agent design
search
minimax + pruning + state cache
utility fn
progress, capture, tempo, risk
weight tuning
Bayesian opt (GP + EI)
chance model
full yut outcome distribution
AI-powered File Organizer
Python
NLP
PyInstaller
Watchdog
spaCy
- Developed an automated file organization system using AI-driven content-based classification, improving file management efficiency by 40%.
- Calculated document similarity using two approaches: TF-IDF vectorization with cosine similarity, and semantic embeddings from spaCy’s en_core_web_md model.
- Integrated real-time file monitoring with Watchdog to automatically organize files into user-defined folders, reducing manual sorting time by 60%.
Agentic Automation · NLP · Embeddings
AI File Organizer
agent pipeline
1
real-time event trigger
Watchdog monitors filesystem; new files emit classification tasks
2
semantic embedding
spaCy en_core_web_md encodes document content into dense vectors
3
similarity classification
Cosine similarity over embeddings; TF-IDF cosine as fallback
4
autonomous dispatch
Agent routes file to closest matching folder without user input
40%
management
efficiency gain
efficiency gain
60%
manual sorting
time saved
time saved
technical design
classification
content-based, not extension
embedding model
spaCy en_core_web_md
fallback
TF-IDF + cosine similarity
deployment
standalone .exe via PyInstaller
Statistical Decision Making
Python
Bayesian Inference
Classification Models
- Implemented Bayesian inference methods (MLE, MAP, posterior mean) for robust parameter estimation in probabilistic models.
- Optimized inventory with the Newsvendor Problem, reducing losses by 49.08% vs heuristic methods.
- Developed and deployed multiple classification models, including KNN, Logistic Regression, and Single feature models, to predict healthcare readmissions, achieving AUCs of 0.68, 0.80, and 0.78, respectively.
- Achieved up to 7.8% cost savings through predictive model optimization, improving decision-making in healthcare resource allocation.
Bayesian · Optimization · Reinforcement Learning
Statistical Decision Making
problems
Newsvendor / inventory optimization
censored demand, weather covariate integration
49% loss reduction vs heuristic
Healthcare readmission prediction
threshold-optimized misclassification cost
AUC 0.80 · 7.8% cost savings
Bayesian parameter estimation
Beta-Binomial conjugate, MLE / MAP / posterior mean
Multiple Secretary Problem
optimal stopping, unknown distribution
tabular Q-learning + ε-greedy
49%
inventory loss
reduction
reduction
0.80
best AUC
readmission
readmission
models
classification
LR · KNN · single-feature
Bayesian
conjugate Beta-Binomial
inventory
newsvendor + demand censoring
stopping
Q-learning, ε-greedy exploration
Real-time Sarcasm Detector
Python
PyTorch
NLP
Built a BERT-based sarcasm classifier with Hugging Face and TweetEval for real-time inference, demonstrating advanced natural language processing capabilities.
View on GitHub