Haneul Kim
Portrait of Haneul Kim

Haneul Kim

B.S. Computer Science, Machine Learning and AI, California Baptist University

Human-AI interaction, AI and LLM evaluation, grounding, synthetic data generation, and AI applications in biomedicine.

I'm a senior at California Baptist University. Most of my research asks whether language models can be trusted to say what their sources actually support. In the CBU AI and ML Lab, I build citation-grounded QA models for U.S. Army doctrine, where I found a model citing sources in every answer while almost never citing the right ones. That gap between what a metric shows and what a model does is the problem I want to keep working on.

I've also applied machine learning to flood prediction and physiological biometrics, and built production LLM systems in industry.

Email me LinkedIn GitHub

Research

Five projects across language models and scientific ML

Doctrine1,073 chunksGenerateGemma 4 26BVerify8 gatesFine-tuneLoRAEvaluate1,071 MCQsfullpartialnoneAnswerability
The doctrine QA pipeline. Orange: the answerability set I designed.

LLM Reliability for Military Doctrine QA

Feb 2026 – Present

ML Researcher, CBU AI and ML Lab (CAiMLL)
Dr. Mark Kim, Dr. Esther Lee, Dr. Shannon Siojo, Dr. Jounsup Park

Training models to answer questions on U.S. Army doctrine with citations that actually support the answer.

  • Grounding
  • LoRA
  • Evaluation
Details
  • Built the data generation, verification, and LoRA training pipeline for a citation-grounded QA model on U.S. Army doctrine.
  • Found the model cited sources in every answer but rarely the right ones (~0.1 F1), and redesigned the dataset to target doctrinal reasoning.
  • Designed answerability-aware augmentation so the model learns when a question can't be answered from its source.
Dot plot of ROUGE-L by question type: fine-tuned Gemma beats the base model on all eight types, average 0.199 vs 0.116
Fine-tuning Gemma on a faculty handbook raised ROUGE-L on every question type (30 validation questions).

Separate vs. Unified Fine-Tuning for Multi-Document Domains

May 2026 – Present

Undergraduate Researcher, with Kip Park, California Baptist University
Dr. Esther Lee, Dr. Shannon Siojo, Dr. Jounsup Park

Should one model learn many documents, or should each document get its own model?

  • Hallucination
  • Fine-tuning
  • Benchmarks
Details
  • Co-led a study testing whether one unified model or one model per document better reduces hallucination in document QA.
  • Fine-tuned Gemma on three universities' handbooks and built standard and near-miss multiple-choice benchmarks.
  • Found the better strategy depends on difficulty: unified won on standard questions, specialists on harder near-miss ones.
Three line charts, one per museum (OCMA, CC, ESC), of mean PWI at six exhibit stops for Catholic, Protestant, and non-religious participants; all groups follow a similar U shape, highest at the first and last stops
Mean PWI at each exhibit stop by religious affiliation, across three museums.

Physiological Biometrics Research

Jan 2026 – Present

Data Science Research Assistant, Templeton Research
Dr. Lee (Stanford University collaboration)

Reading museum visitors' physiological responses from wearable sensors.

  • Wearables
  • Eye tracking
  • Signal processing
Details
  • Analyze museum visitors' physiological responses from wearable sensors in a study with Stanford.
  • Fixed timestamp misalignment between biometric and eye-tracking devices and built a sensor gap-detection tool.
Observed mountain gauge rainfall with LSTM ensemble predictions closely tracking it; test set R squared 0.766
LSTM ensemble predictions against the observed mountain gauge during test storms.

Hydrology and Flood Prediction

Sep 2025 – Present

Undergraduate Researcher, California Baptist University
Dr. Benjamin Sanders

Filling gaps in a mountain rain gauge to feed flood modeling for the City Creek watershed.

  • LSTM
  • PyTorch
  • Time series
Details
  • Built the ML pipeline filling gaps in a mountain rain gauge from a lower-elevation gauge, feeding flood modeling for the City Creek watershed.
  • Built LSTM and CNN-LSTM models in PyTorch; LSTM reached mean R² = 0.758 ± 0.024 over 15 seeds, outperforming CNN-LSTM on R², MAE, and KGE.
  • Replaced a noisy single-split early-stopping scheme with 5-fold chronological cross-validation, stabilizing training across seeds.
  • Benchmarked against 11 regression models and 5 MLPs; a 15-seed LSTM ensemble (R² = 0.766) slightly exceeded the linear baseline (R² = 0.764).
Light rays deflected by many masses, rendered in Manim.

Gravitational Lensing Simulation

Jan 2025 – Jul 2025

Physics Research Assistant, California Baptist University
Dr. Hyung Choi

Computational simulation of generalized gravitational lensing, presented at the CBU research symposium.

  • Simulation
  • Manim

Publications and presentations

  1. AbstractIGARSS 2026

    Towards Modelling Runoff and Flood Prediction over Unreliable Precipitation Gauges and Assessment of Potential Gauge Error

    Jung, H., Sanders, B., Kim, H., & Park, K.

    IEEE International Geoscience and Remote Sensing Symposium (IGARSS)

    AcceptedFull paper in preparation, Jan 2027
  2. PosterSCCUR 2026

    Separate or Unified Fine-Tuning? A Systematic Comparison of Training Strategies for Multi-Document Domain Applications

    Kim, H., & Park, K.

    Southern California Conferences for Undergraduate Research, San Diego State University

    Nov 21, 2026
  3. PosterSCCUR 2026

    Towards Modelling Runoff and Flood Prediction over Unreliable Precipitation Gauges

    Park, K., Kim, H., Jung, H., & Sanders, B.

    SCCUR, San Diego State University

    Nov 21, 2026
  4. TalkCBU 2026

    Computational Simulation of Generalized Gravitational Lensing

    Kim, H.

    18th Annual Natural and Mathematical Sciences Research Symposium, California Baptist University

    Mar 28, 2026
  5. PosterKSEA 2026

    Towards a High Accuracy Earthquake Forecasting Mobile App

    Kim, H., & Sanders, B.

    33rd Annual KSEA South-Western Regional Conference, California State University, Fullerton

    Jan 31, 2026

News

  1. Started as TA for PHYS 2140, Physics for Scientists I.

  2. Abstract accepted at IEEE IGARSS.

  3. Received the KSEA Undergraduate Scholarship.

  4. Talk on gravitational lensing simulation at the CBU research symposium.

  5. Poster on earthquake forecasting at the KSEA South-Western Regional Conference.

Industry

AI and Data Automation Intern / IT Assistant

Jun – Sep 2026

Elim, Inc., Irvine, CA

  • Deployed a two-node NVIDIA DGX Spark cluster (200 Gb/s RDMA) and staged a 284B-parameter MoE model for local inference for automation tasks.
  • Shipped a multi-agent LLM app (6 debating agents plus moderator) with output validators and blind-evaluation benchmarking; 1,480+ tests.

AI Engineering Intern, Lance Chatbot

Jan – Jul 2026

CBU Campus Store

  • Built and deployed a production RAG chatbot resolving textbook access issues across 11+ publisher platforms.
  • Ran fully on-premise (FastAPI, FAISS, Llama 3.2 via Ollama) for FERPA compliance, with sub-50ms retrieval.
  • Added similarity-threshold gating to block out-of-scope hallucinations, plus request queuing to prevent GPU saturation.
Lance chatbot answering a McGraw Hill Connect access question, with recommended PDF guides in a side panel

Selected projects

BAY-min Personal Tutor

Sep – Dec 2026

A voice-interactive personal tutor on the Reachy Mini robot, built with the Gemini Live API.

A fixed-wing drone model selected in the NVIDIA Isaac Sim viewport, casting a shadow on a gray ground plane

HAPS (Hostile Aircraft Protection System)

2026 – 2027, Senior Capstone, NSWC Corona

Counter-UAS software (detection, tracking, sensor fusion, operator app, simulation) on a three-person team.

Forward passLinear1 → 32ReLUor GELULinear32 → 1MSE lossvs. noisy sin(x)Backpropagation: loss.backward()SGD updateweights and biases, every epoch

GUTT, Large Language Model from Scratch

May – Aug 2025, ACM Software Engineering Team

A transformer built from scratch, starting with a feedforward network and hand-implemented activations.

TA

  • TA, Physics for Scientists I (PHYS 2140)Fall 2026 California Baptist University, Dr. Hyung Choi

    Support 70+ science and math majors in calculus-based mechanics, thermodynamics, fluids, and waves.

  • TA and Mentor, 3 Pasos ProgramSummer 2025 California Baptist University, NSF Grant 22-611, HSI Initiative

    Designed STEM labs and mentored high school students in mathematical reasoning.

  • TA, Mathematics for Engineering (EGR 181)Fall 2024 California Baptist University

    Supported ~90 students applying calculus, linear algebra, and differential equations to engineering problems; used MATLAB for simulations and assisted in circuit building labs.

Honors and skills

  • KSEA Undergraduate ScholarshipKorean-American Scientists and Engineers Association, 2026
  • Alpha Chi National Honors SocietyCalifornia Baptist University

Programming

  • Python
  • C++
  • Java
  • R
  • SQL
  • JavaScript
  • Dart

ML and data

  • PyTorch
  • scikit-learn
  • XGBoost
  • LoRA
  • DeepSpeed
  • Unsloth
  • FAISS
  • Ollama
  • NumPy
  • Pandas

Systems and tools

  • Docker
  • Linux
  • Git
  • NVIDIA DGX
  • GCP
  • BigQuery
  • FastAPI
  • React
  • Firebase
  • MATLAB
  • Manim

Languages

  • English
  • Korean
  • Spanish