KSTAR · Seoul National University
KSTAR Plasma Density Forecasting
An end-to-end forecasting pipeline for KSTAR tokamak plasma density, built during a research internship at Seoul National University (May–Aug 2026).
Overview
Built an end-to-end forecasting pipeline for 157,000+ KSTAR tokamak sensor files across 1,100+ plasma shots — covering extraction, shot-level splitting, normalization, modeling, and long-horizon evaluation. Diagnosed a state-reset flaw in a windowed LSTM baseline and developed a stateful model that preserves memory during continuous autoregressive rollout.
Problem
Note: specific framing of what made plasma density forecasting difficult (e.g. shot-to-shot variability, sensor noise, horizon length) — to be filled in later.
Approach
Experiments / Iteration
Results
Lessons / Takeaways
Note: personal reflection (what worked, what didn't, what you'd try next) — to be filled in later.
Technologies
Python · PyTorch · NumPy · Pandas
KAIST · Transportation Risk Mapping
Multi-Layer Transportation Risk Mapping with LLM-Based Explanation
A two-stage spatial ML pipeline for intersection-level crash-risk prediction, paired with a retrieval-grounded explanation layer — developed with KAIST and Texas A&M University (Jun–Jul 2026).
Overview
Co-developed a two-stage spatial ML pipeline with Texas A&M University for intersection-level crash-risk prediction, then built a retrieval-grounded explanation layer on top so predictions come with human-readable, evidence-based justifications rather than opaque scores. Co-first author on the resulting manuscript, submitted to the Transportation Research Board (TRB) in August 2026.
Problem
Note: specific framing of the crash-risk prediction task and why it was difficult — to be filled in later.
Approach
Experiments / Iteration
Results
Crash-risk prediction (supporting result — the pipeline this explanation layer sits on top of):
Lessons / Takeaways
Note: personal reflection (what worked, what didn't, what you'd try next) — to be filled in later.
Technologies
Python · OpenAI API · RAG / retrieval grounding · TreeSHAP · Pandas
KNUE · LLM Research
LLM Evaluation of K–12 Science Classroom Dialogue
Testing retrieval and prompting strategies for reliable LLM evaluation of K–12 science classroom dialogue — LLM Researcher, Korea National University of Education (Jan–May 2026, remote).
Overview
Evaluated LLM-based methods for analyzing K–12 science classroom dialogue at the Korea National University of Education. The system scored individual utterances on a 1–5 scale with justification and classified dialogue into 20 instructional themes. I systematically tested retrieval-based demonstration selection, contextual augmentation, and reasoning strategies to understand how prompt construction and example selection affected model behavior.
Problem
Classroom utterances are often short and context-dependent — individual responses can be ambiguous without the surrounding dialogue. The rating distribution was also highly imbalanced (rating 1 ≈ 0.4%, rating 2 ≈ 3.0%, rating 3 ≈ 85.5%, rating 4 ≈ 10.7%, rating 5 ≈ 0.4%), so naive few-shot demonstration selection risked overrepresenting the dominant class. Structured educational evaluation requires consistency, not just fluent-sounding output.
Approach
Semantic retrieval used multilingual-e5-base embeddings to find relevant demonstrations. Rating-stratified demonstration banks kept example selection from being dominated by the majority score. Contextual augmentation tested previous and next dialogue turns to disambiguate short utterances. Prompting/reasoning strategies were evaluated rather than assumed to help.
Experiments / Iteration
Research questions tested: does semantic similarity improve demonstration selection over fixed or naive examples? Does rating-stratified retrieval improve representation of rare scores? Does surrounding dialogue improve interpretation of ambiguous utterances? Does additional reasoning improve the final rating? Which strategies improve reliability, and which introduce new failure modes?
Findings & Evaluation
The experiments were most useful for identifying how retrieval, context, and prompt construction changed model behavior, and where those approaches still failed under the rating task.
Semantic and rating-aware demonstration selection provided a more controlled way to construct few-shot prompts than relying on naive or fixed example selection.
Previous and next utterances provided useful information for short responses that were difficult to interpret in isolation, while also introducing a tradeoff between additional context and prompt complexity.
Reasoning and justification strategies were evaluated experimentally rather than treated as inherently beneficial, helping identify where additional reasoning changed or failed to improve final judgments.
Cross-validation and error analysis showed that reliable minority-class prediction remained difficult, highlighting limitations of LLM-based scoring under heavily imbalanced educational data.
- 5-fold CV accuracy: 0.5449 ± 0.0064
- Macro-F1: ≈ 0.20
Included here for transparency as part of the evaluation, not as the main accomplishment of the project — minority-class prediction remained difficult under this level of class imbalance.
Lessons / Takeaways
- Few-shot prompting depends heavily on which demonstrations are selected
- Retrieval design matters when the label distribution is highly skewed
- More conversational context can help resolve ambiguity but is not automatically beneficial
- Additional LLM reasoning should be validated experimentally rather than assumed to improve results
- Aggregate accuracy is not sufficient for understanding model behavior under class imbalance
- Error analysis and controlled comparisons were essential for identifying where the system remained unreliable
Technologies
Python · OpenAI API · multilingual-e5-base · Embedding Retrieval · NLP · Pandas · scikit-learn