SYS_SHEET_01 / COVER

Portfolio for Technical Review

Data & Applied AI Engineer

I work across data structure, AI development workflows, and NLP/LLM evaluation, carrying problems from framing to implementation, validation, and delivery.

[SYS_ROLE: EXECUTIVE_SUMMARY] · Role Summary

Data Systems Data Systems Specialist with Graph-native Depth

Turns heterogeneous data into model-ready and system-ready structures, pipelines, and graph representations.

AI-DLC / MLOps AI-DLC and Operational MLOps Engineer

Makes run records, artifacts, model behavior, and feedback paths inspectable enough to improve.

NLP / LLM Applied NLP and LLM Research Engineer

Translates language-model research into better modeling and evaluation decisions in applied systems.

[SYS_EVIDENCE: VERIFICATION_MATRIX]

Case Study / Evidence	Data Systems	AI-DLC / MLOps	NLP / LLM
Featured Case Studies
2026 / Self-directed ML Platform Project Contexta: Local-First ML Observability	Supporting	Primary	—
2026 / Graph Systems Engine Project Lynxes: Graph Analytics Engine	Primary	Supporting	—
2025 / Clinical AI Research EMR-Based Nursing Surveillance for Automatic ICD Coding	Supporting	—	Primary
2024 / Internal Commerce Platform Dalkom Shop: Internal Employee Mileage Commerce Platform	Supporting	Primary	—
2024 / LLM Application Prototype Devridge: LLM-Based Feedback Bridge for Developers	—	Supporting	Primary
2023 / Collaborative NLP Project BloGeek: AI Modules for a React + Spring Blog Project	—	Supporting	Primary
2023 / Conversational AI Project FRIMO: Conversational AI for Emotional Support and Diary Generation	—	Supporting	Primary
Research Publications
2026 / Research A Context-Adaptive Gated Embedding Framework for Advanced Clinical Decision-Making	Supporting	—	Primary
2025 / Research Deep Learning based Automatic ICD Coding for Nursing Surveillance of Abdominal Surgery Patients	Supporting	—	Primary
2025 / Research Empathetic Dialogue Generation Model Using Reinforcement Learning with AI-Based Feedback	—	—	Primary
Capability Study Groups
2023 / Study NEKA: NLP 스터디 (KoBERT)	—	—	Primary
2023 / Study Sunaroum: Financial AI Study	—	Supporting	Primary
2022 / Study Ida: Recommender Systems Study	Primary	—	—
2022 / Study MLADS: 머신러닝 기초 스터디	Primary	Supporting	—

SYS_SHEET_02 / CASE_STUDY

[SYS_PIPELINE: TIMELINE]

WARN Problem

ML experiments and deployment work often scatter metadata, records, and artifacts across tools, making reproducible local observability hard to maintain.

EXEC Key Decision

Used a `.contexta/` workspace as the home for separated metadata, records, and artifact storage.

OK Outcome

Implemented a local observability foundation for consistently managing and inspecting ML execution history and artifacts.

[SYS_FLOW: ARCHITECTURE]

A local-first ML observability structure where the Python facade, CLI, workspace storage, and report/recovery flows share one contract.

Python API → CLI → Canonical Contract → Local Workspace → Reports & Recovery

[SYS_SPEC: CONSTRAINTS] · Constraints

The system had to work from a local workspace without an external backend.
The library API and CLI needed to share the same contract.
Execution records, artifacts, comparison reports, and recovery flows had to feel like one product.

[SYS_DECISIONS: DESIGN_TRADE_OFFS] · Design Decisions

EXEC

Canonical workspace

Used a `.contexta/` workspace as the home for separated metadata, records, and artifact storage.

Rationale: This made execution history inspectable through ordinary local files.

EXEC

Schema-first contract

Made capture, query, compare, and recovery flows share the same data meaning.

Rationale: The contract keeps future features from drifting into incompatible data shapes.

This project reinforced that ML tooling becomes product-like when execution traces remain explainable after the run finishes.

[SYS_ROLE: TELEMETRY]

Supporting Data Systems

Primary Focus AI-DLC / MLOps

— NLP / LLM

[SYS_METRIC: IMPACT] · Review Points

Local-first · Architecture

Schema-first · Contract

Trace / Compare / Recover · Workflow

Shows operational observability design for tracing, comparing, and recovering AI execution records and artifacts.

[SYS_SPEC: METADATA]

Data Surfaces

structuredhybrid

AI-DLC Stages

dataexperimentevaluationobservabilityfeedback-recovery

SYS_SHEET_03 / CASE_STUDY

[SYS_PIPELINE: TIMELINE]

WARN Problem

Existing Python graph libraries and generic dataframe wrappers often struggle to combine memory efficiency, traversal performance, and lazy query optimization for large graph analytics.

EXEC Key Decision

Designed GraphFrame to own Arrow RecordBatches directly.

OK Outcome

Established the foundation for a graph analytics engine with Arrow columnar memory, CSR-based traversal, and lazy collect execution.

[SYS_FLOW: ARCHITECTURE]

A graph analytics engine combining Arrow RecordBatch storage, CSR indexing, lazy logical plans, Rust algorithms, and Python bindings.

Arrow RecordBatch → GraphFrame → CSR Index → LogicalPlan → Python Bindings

[SYS_SPEC: CONSTRAINTS] · Constraints

Node and edge data needed a column-oriented memory layout.
Neighbor traversal had to avoid linear scans.
The Rust engine and Python usability layer needed a stable boundary.

[SYS_DECISIONS: DESIGN_TRADE_OFFS] · Design Decisions

EXEC

Arrow-owned GraphFrame

Designed GraphFrame to own Arrow RecordBatches directly.

Rationale: This combines columnar memory efficiency with compatibility potential for Python analytics workflows.

EXEC

CSR adjacency index

Added CSR-based adjacency indexing to EdgeFrame so neighbor lookup becomes O(degree).

Rationale: Traversal should behave like a graph-native operation rather than a generic table scan.

EXEC

Lazy logical plan

Accumulated queries into a LogicalPlan and deferred execution until `.collect()`.

Rationale: This keeps the analysis pipeline optimizable.

The project clarified that system performance becomes convincing when data ownership, memory layout, and user API design work together.

[SYS_ROLE: TELEMETRY]

Primary Focus Data Systems

Supporting AI-DLC / MLOps

— NLP / LLM

[SYS_METRIC: IMPACT] · Review Points

O(degree) · Neighbor lookup

Arrow · Memory model

Lazy collect · Execution

Shows system design and implementation depth for treating graph data as a first-class execution model.

[SYS_SPEC: METADATA]

Data Surfaces

structuredgraphhybrid

AI-DLC Stages

dataexperimentevaluation

SYS_SHEET_04 / CASE_STUDY

[SYS_PIPELINE: TIMELINE]

WARN Problem

Nursing surveillance required diagnosis-related classification, but key clinical signals were fragmented across heterogeneous EMR sources.

EXEC Key Decision

Trained two KM-BERT models independently and averaged raw logits to stabilize text representation.

OK Outcome

The final model was reviewed for practical classification behavior rather than a single standalone score, including rare-class recall and available-data constraints.

[SYS_FLOW: ARCHITECTURE]

A clinical AI pipeline that processes structured EMR and nursing text in parallel before stacking them for ICD prediction.

Structured EMR → Nursing Text → Dual KM-BERT → PCA + XGBoost → Rare-class Evaluation

[SYS_SPEC: CONSTRAINTS] · Constraints

The model could not depend on discharge summaries or other post-event documents.
Rare classes still needed practically meaningful recall.
The pipeline had to combine structured EMR features with Korean clinical text.

[SYS_DECISIONS: DESIGN_TRADE_OFFS] · Design Decisions

EXEC

Dual KM-BERT representation

Trained two KM-BERT models independently and averaged raw logits to stabilize text representation.

Rationale: The ensemble reduced volatility in clinical text signals.

EXEC

PCA and XGBoost stacking

Reduced BERT-derived representations with PCA and used XGBoost as the final ICD classifier.

Rationale: This made high-dimensional text signals and structured EMR features easier to combine.

The work clarified that clinical AI value depends not only on performance, but on which documents and signals are realistically available at the time of use.

[SYS_ROLE: TELEMETRY]

Supporting Data Systems

— AI-DLC / MLOps

Primary Focus NLP / LLM

[SYS_METRIC: IMPACT] · Review Points

Balanced · Evaluation Scope

Core EMR · Available Data

High · Rare-class Recall

Combines heterogeneous structured EMR and Korean clinical text into an evaluable NLP pipeline.

[SYS_SPEC: METADATA]

Data Surfaces

structuredtexthybrid

AI-DLC Stages

datatrainingevaluation

[SYS_LINK: ARTIFACTS]

Published paper Paper

SYS_SHEET_05 / CASE_STUDY

[SYS_PIPELINE: TIMELINE]

WARN Problem

The project needed cloud infrastructure and platform foundations that could reliably support search, notifications, admin workflows, and ongoing operations.

EXEC Key Decision

Organized CI/CD and cloud foundations so the React and Spring service could fit into a deployable, monitorable workflow.

OK Outcome

Built a practical service foundation for a closed internal mileage commerce platform.

[SYS_DECISIONS: DESIGN_TRADE_OFFS] · Design Decisions

EXEC

Platform foundation

Organized CI/CD and cloud foundations so the React and Spring service could fit into a deployable, monitorable workflow.

[SYS_ROLE: TELEMETRY]

Supporting Data Systems

Primary Focus AI-DLC / MLOps

— NLP / LLM

[SYS_METRIC: IMPACT] · Review Points

CI/CD + Cloud · Foundation

Shows delivery, security, and observability foundations for running service features in an operational environment.

[SYS_SPEC: METADATA]

Data Surfaces

structuredimagehybrid

AI-DLC Stages

datadeploymentobservabilityfeedback-recovery

SYS_SHEET_06 / CASE_STUDY

[SYS_PIPELINE: TIMELINE]

WARN Problem

Developers working alone often need UI, performance, or code quality feedback, but they rarely have an easy way to gather role-specific input at the right time.

EXEC Key Decision

Separated role constraints and contextual input so responses stayed scoped instead of generic.

OK Outcome

The prototype showed how structured prompting can make LLM feedback more useful for project review.

[SYS_DECISIONS: DESIGN_TRADE_OFFS] · Design Decisions

EXEC

Role-aware prompting

Separated role constraints and contextual input so responses stayed scoped instead of generic.

[SYS_ROLE: TELEMETRY]

— Data Systems

Supporting AI-DLC / MLOps

Primary Focus NLP / LLM

[SYS_METRIC: IMPACT] · Review Points

Role-based · Feedback modes

Shows prompt and interaction design for making LLM output useful in role-based technical review.

[SYS_SPEC: METADATA]

Data Surfaces

text

AI-DLC Stages

dataexperimentinferenceevaluation

SYS_SHEET_07 / CASE_STUDY

[SYS_PIPELINE: TIMELINE]

WARN Problem

The product needed ML components that could classify emotional polarity and generate stylistic variations of text to support richer blog content workflows.

EXEC Key Decision

Used KoBERT for polarity recognition and KoBART for style transfer.

OK Outcome

The project gave the team practical AI modules for blog-oriented text processing.

[SYS_DECISIONS: DESIGN_TRADE_OFFS] · Design Decisions

EXEC

Separated KoBERT and KoBART roles

Used KoBERT for polarity recognition and KoBART for style transfer.

[SYS_ROLE: TELEMETRY]

— Data Systems

Supporting AI-DLC / MLOps

Primary Focus NLP / LLM

[SYS_METRIC: IMPACT] · Review Points

Classification + Generation · Model scope

Connects Korean NLP classification and generation models to product-shaped web service features.

[SYS_SPEC: METADATA]

Data Surfaces

text

AI-DLC Stages

dataexperimenttraininginferenceevaluation

SYS_SHEET_08 / CASE_STUDY

[SYS_PIPELINE: TIMELINE]

WARN Problem

The product needed ML components that could recognize user emotion and support a diary-generation workflow from daily conversation logs.

EXEC Key Decision

Centered the ML workflow around KoBERT-based emotion recognition and connected it to chatbot and summarization flows.

OK Outcome

The project delivered an MVP-level conversational diary experience with an emotion recognition pipeline.

[SYS_DECISIONS: DESIGN_TRADE_OFFS] · Design Decisions

EXEC

Emotion-first ML pipeline

Centered the ML workflow around KoBERT-based emotion recognition and connected it to chatbot and summarization flows.

[SYS_ROLE: TELEMETRY]

— Data Systems

Supporting AI-DLC / MLOps

Primary Focus NLP / LLM

[SYS_METRIC: IMPACT] · Review Points

Korean NLP · Pipeline

Connects Korean NLP models into a user-facing AI pipeline for conversational product experience.

[SYS_SPEC: METADATA]

Data Surfaces

text

AI-DLC Stages

dataexperimenttraininginferenceevaluation

SYS_SHEET_09 / RESEARCH_CASE

[SYS_RESEARCH: ABSTRACT]

This study proposes a hierarchical clinical decision support framework that estimates diagnostic context via partial-label automated ICD coding and reinjects it into irregular ICU time-series forecasting through context-adaptive gating for mechanical ventilation transition prediction. By conditioning temporal interpretation on diagnostic context, the framework substantially improves rare-event detection.

[SYS_RESEARCH: BACKGROUND] · 연구 배경

This research was designed to overcome key bottlenecks and academic limits in the domain. It addresses EMR structured data and unstructured text/embedding domain complexities, showing modeling judgment that drives real-world value.

[SYS_RESEARCH: METHODOLOGY] · 핵심 기여

Academic Claim: Proposed a context-adaptive gated embedding framework that reinjects diagnostic context from automated ICD coding into ICU time-series prediction.

Methodology: Combines partial-label diagnostic context, TCN temporal representations, and gating to strengthen rare transition-event prediction.

[SYS_ROLE: TELEMETRY]

Primary Focus NLP / LLM

Supporting Data Systems

[SYS_METRIC: RELEVANCE]

Shows modeling judgment that connects diagnostic context and time-series signals hierarchically instead of treating data surfaces as isolated inputs.

[SYS_RESEARCH: BIBTEX]

@article{kim2026cage,
  title={A Context-Adaptive Gated Embedding Framework for Advanced Clinical Decision-Making},
  author={Donghyeon Kim and Daeho Kim and Okran Jeong},
  journal={Mathematics},
  year={2026},
  note={{Submitted / under review}}
}

[SYS_SPEC: METADATA]

Clinical Decision Support SystemAutomated ICD CodingICU Time-seriesMechanical Ventilation PredictionPartial-Label LearningExtreme Multi-Class ClassificationTCNGatingRare Event Detection

SYS_SHEET_010 / RESEARCH_CASE

[SYS_RESEARCH: ABSTRACT]

This study proposes an automatic ICD coding model for nursing surveillance of abdominal surgery patients by integrating EMR-based test data, patient information, and nursing notes. A stacking architecture combining dual KM-BERT, XGBoost, and PCA outperformed both a single KM-BERT model and simpler ensemble variants.

[SYS_RESEARCH: BACKGROUND] · 연구 배경

This research was designed to overcome key bottlenecks and academic limits in the domain. It addresses EMR structured data and unstructured text/embedding domain complexities, showing modeling judgment that drives real-world value.

[SYS_RESEARCH: METHODOLOGY] · 핵심 기여

Academic Claim: Proposed a deep-learning ICD coding model that combines structured EMR information and nursing notes for nursing surveillance.

Methodology: Used dual KM-BERT, PCA, and XGBoost stacking to evaluate sparse and domain-sensitive diagnosis labels.

[SYS_ROLE: TELEMETRY]

Primary Focus NLP / LLM

Supporting Data Systems

[SYS_METRIC: RELEVANCE]

Supports the portfolio claim that NLP/LLM systems should be judged through domain data structure and error distribution, not only headline accuracy.

[SYS_RESEARCH: BIBTEX]

@article{kim2025deep,
  title={Deep Learning based Automatic ICD Coding for Nursing Surveillance of Abdominal Surgery Patients},
  author={Donghyeon Kim, Daeho Kim, Seyoung Kim, Okran Jeong},
  journal={Journal of The Korea Society of Computer and Information},
  volume={30},
  number={5},
  pages={21--30},
  year={2025},
  publisher={The Korean Society Of Computer And Information}
}

[SYS_SPEC: METADATA]

Medical AINursing SurveillanceEMRAutomatic ICD CodingDeep LearningKM-BERTXGBoostEnsembleAbdominal Surgery

SYS_SHEET_011 / RESEARCH_CASE

[SYS_RESEARCH: ABSTRACT]

This study proposes an empathetic dialogue generation model using reinforcement learning with AI-based feedback (RLAIF) to address limited diversity and reliance on human feedback. By leveraging an LLM as a reward evaluator and integrating it into EmpRL, the model generates more diverse empathetic responses.

[SYS_RESEARCH: BACKGROUND] · 연구 배경

This research was designed to overcome key bottlenecks and academic limits in the domain. It addresses EMR structured data and unstructured text/embedding domain complexities, showing modeling judgment that drives real-world value.

[SYS_RESEARCH: METHODOLOGY] · 핵심 기여

Academic Claim: Proposed an RLAIF-based reinforcement learning structure to reduce reliance on human feedback in empathetic dialogue generation.

Methodology: Used an LLM evaluator as a reward signal and PPO-based policy updates to improve response diversity and empathy alignment.

[SYS_ROLE: TELEMETRY]

Primary Focus NLP / LLM

[SYS_METRIC: RELEVANCE]

Supports the view that LLM response quality should be treated as an evaluation, reward, and policy-update system rather than a prompt-only outcome.

[SYS_RESEARCH: BIBTEX]

@inproceedings{joo2025empathetic,
  title={Empathetic Dialogue Generation Model Using Reinforcement Learning with AI-Based Feedback},
  author={Yongwan Joo, Donghyun Lim, Donghyeon Kim, Seungyeon Sun, Okran Jeong},
  booktitle={Proceedings of the Korea Computer Congress 2025},
  pages={2410--2412},
  year={2025}
}

[SYS_SPEC: METADATA]

Empathetic DialogueReinforcement LearningRLAIFRLHF AlternativeLLMDialogue GenerationNLPAI Feedback

SYS_SHEET_12 / CREDENTIALS_GROWTH

[SYS_CREDENTIAL: PATHWAY] · 자격 및 학습 경로

// 01_ACADEMIC_STUDIES

2024.03 - 2026.02

Gachon University

M.S. in Artificial Intelligence · GPA 4.14/4.5. Intelligent Data Analytics Lab. Advisor - OkRan Jeong. Focused on clinical AI, evaluation, and implementation-oriented research delivery.

FOCUS: Advanced ML, Graph Neural Networks (GNN), Medical Time-Series

2019.03 - 2024.02

Gachon University

B.S. in Software · Gained leadership and community management experience by participating in an official programming club and leading various study groups, including basic ML, advanced ML, financial ML, and GNN.

FOCUS: Data Structures, Algorithms, Linear Algebra, Statistical ML

// 02_PROFESSIONAL_COURSES

2023.11

Practical Implementation of Monitoring and Testing in DevOps Environments

LLOYDK · Completed practical training in Elastic-based DevOps monitoring and testing.

2023.12

Multi Cloud Orchestration Program

5Works · Completed HashiCorp-based multi-cloud orchestration and IaC training.

2024.02

Company-Led Intensive Project Training

DK Techin · Participated in industry-linked practical training focused on security and DevOps engineering.

2024.02

Micro Degree in Software Specialist Training

Gachon University · Completed a micro-degree program for training software specialists.

// 03_CONTINUOUS_LEARNING

2024.03

AWS Builders Online Series

AWS · Learned serverless architectures and generative AI application integration workflows based on Amazon Bedrock

2023.08

Open-source NLP/LLM Continuous Learning

Community · Analyzed open-source LLM parameter-efficient fine-tuning (LoRA, QLoRA) and RAG pattern architectures

[SYS_LOG: PIPELINE_EXECUTION] · 시스템 훈련 로그

[INFO] 2026-05-25 02:09:03 - Initializing training pipeline...

[SPEC] CUDA Version: 12.2 / Device: NVIDIA RTX 4090 x2

[LOAD] Loading pretrained model: kobert-base-v3-medical
· Config: 12-layer, 768-hidden, 12-heads (110M Params)

[TRAIN] Epoch [4/10] - Step [4500/12000] - Loss: 0.1842
· Optimizer: AdamW (lr=3e-5) with linear warmups

[EVAL] Evaluating tokenization on clinical NER testsets...
· F1-Score: 0.8816 / Precision: 0.8842 / Recall: 0.8791

* SYSTEM SIMULATION LOG FOR VISUALIZATION PURPOSE.
* ALL SOURCE ALGORITHMS MAPPED TO LIVE PRODUCTION SERVICES.

[SYS_CREDENTIAL: STUDY_GROUPS] · 자기주도 역량 개발 스터디

NEKA

NLP Study

A study on implementing Keyword Extraction and NER with KoBERT

STACK: PyTorch, Transformers, KoBERT, Scikit-Learn

OBJECTIVE:

Advanced training on morphosemantics, keyword extraction, and named entity recognition (NER) using KoBERT in specialized medical and native language domains

KEY OUTCOMES:

Designed and implemented a custom KoBERT fine-tuning pipeline in PyTorch, achieving 0.88 NER F1-Score on medical clinical records and validation datasets

LESSONS LEARNED:

"Analyzed Multi-Head Attention maps to study medical token boundary segmentation, proving the high sensitivity of specialized domain vocabulary embeddings on boundary accuracy"

RESEARCH REFERENCE:

Theoretical analysis of Transformer architecture in Attention Is All You Need (Vaswani, 2017) and BERT (Devlin, 2018)

SUNAROUM

Financial AI Study

A study on building buy recommendation models from financial statements and indicators

STACK:

OBJECTIVE:

KEY OUTCOMES:

LESSONS LEARNED:

""

RESEARCH REFERENCE:

I-DA

Recommender Systems Study

A study on recommendation systems using knowledge graphs and GNNs

STACK:

OBJECTIVE:

KEY OUTCOMES:

LESSONS LEARNED:

""

RESEARCH REFERENCE:

MLADS

Basic Machine Learning Study

A study on the fundamentals of data science and machine learning

STACK: Python, NumPy, SciPy, Matplotlib

OBJECTIVE:

Deep mathematical formulation of machine learning optimization, convergence, and complex regularization mechanisms on top of probability and linear algebra

KEY OUTCOMES:

Derived mathematical equations and implemented dimensionality reductions (PCA, t-SNE) and classifiers from pure NumPy scratch, validating optimization loss

LESSONS LEARNED:

"Acquired precise mental models of convergence bottlenecks, vanishing gradients, and local minima boundary conditions by implementing algorithms mathematically from scratch"

RESEARCH REFERENCE:

Foundations of probability and linear algebra in Deep Learning (Goodfellow, 2016) & PRML (Bishop, 2006) reference studies