CV ↓

About Kip

Kip sitting on a rocky mountain ridge on a clear winter day

hello, world — I’m

Ki-Pung (Kip) Park

AI / ML researcher & engineer · California Baptist University

I’m a Computer Science student at California Baptist University (ML & AI concentration, class of 2027). I work on LLM fine-tuning and evaluation, lead research on subliminal learning, and build AI tools for students and advisors.

Research CVEmailGitHubLinkedIn

I build and evaluate language models: fine-tuning pipelines, careful evaluations, and local-first AI tools that people actually use.

Studying
B.S. Computer Science — Machine Learning & AI concentration
Minor in Christian Studies · Alpha Chi Honor Society
Graduating
Expected Apr. 2027 · GPA 3.94
Interests
LLM architecturesVision-language-action modelsPhysical AI

A bit more about me

I’m majoring in Computer Science with a concentration in Machine Learning & AI, and minoring in Christian Studies. I’m drawn to the place where technology, learning, and creativity meet, and I like sharing what I learn along the way.

As president of CBU’s ACM chapter (2024–2026) I started the Software Engineering Team, where 20+ students built projects like our chapter website. I also tutor software engineering, solve ICPC problems, and enjoy building visual simulations.

When I’m not doing anything “nerdy,” you’ll find me playing basketball with friends, sharing a meal, or hunting for a good series to watch.

Living Kanto.mp4

100 residents, each driven by a local language model, living in one persistent world. Code on GitHub →
Recorded replay. Pokémon graphics belong to their owners.

kip@kipos: ~

KipOS 1.0 · type help to look around.

Research

  1. Feb. 2026 – Present

    LLM Fine-Tuning and Evaluation

    CBU AI & Machine Learning Lab (CAIMLL) · Riverside, CA

    • Built the backbone of the lab’s fine-tuning pipeline: complex reference documents become source-verified training data, with checks for evidence coverage, answer quality, and reproducibility.
    • Designed training examples that separate fully answerable, partially answerable, and unsupported questions, so models learn to admit knowledge gaps while answering what they can.
    • Developed distributed adapter and full-parameter fine-tuning workflows, and audited synthetic data for wording and topic shortcuts.

  2. Aug. 2025 – Present

    Subliminal Learning in LLMs — Study Lead

    California Baptist University · Riverside, CA

    • Led a controlled study comparing LoRA-induced and system-prompt-induced teacher preferences, fine-tuning Llama 3.2 3B students on filtered numerical sequences.
    • Neither pipeline produced robust preference transfer under the tested conditions; leading the manuscript through revision.

  3. Jan. 2026 – Present

    Hydrology — Statistical Rainfall Estimation

    Research collaboration with Dr. Benjamin Sanders

    • Modeled hourly rainfall transfer from CSUSB to Panorama Point using 1996–2024 gauge records; event-calibrated lagged OLS reached R² = 0.58 on held-out observed event hours (≥ 0.04 in/hr).

  4. Jul. – Oct. 2025

    AI Researcher — Mechanistic Interpretability

    Algoverse AI Research Program · Remote

    • Implemented JAX/Haiku transformers for RASP-style algorithmic tasks and sparse-autoencoder pipelines to analyze attention and MLP representations.
    Code on GitHub →

  5. Mar. 2024 – Sept. 2025

    Student Researcher — GeoAI for Public Health

    California Baptist University · Riverside, CA

    • Trained a deep neural network to predict geographic areas of suspected drug-overdose risk, and built the GIS visualizations in Kepler.gl that turned its output into insights; coauthored the AMCIS 2025 paper.

  6. May 2024 – Aug. 2025

    Undergraduate Research Assistant — Computational Physics

    California Baptist University · Adviser: Dr. Hyung Choi · Riverside, CA

    • Built modular, object-oriented simulations of gravitational lensing and interactive visualizations of multi-body lensing behavior.

Papers & talks

Papers

  • Published

    A DSR Approach for Geospatial AI: Prediction of Drug Overdoses.

    Corso, A., Asariah, J. S., Park, K., Corso, N. A., and Lee, E. (2025). Americas Conference on Information Systems (AMCIS), Montréal.

  • Under review

    Replicating Subliminal Learning in LLMs via Distillation and LoRA Fine-Tuning.

    Park, K., Bergquist, G., Lee, P., and Sanders, B. Aletheia — The Alpha Chi Journal of Undergraduate Scholarship.

Presentations

  • Nov. 2026 SCCUR 2026 · San Diego State University

    • Towards Modeling Runoff and Flood Prediction over Unreliable Precipitation Gauges and Assessment of Potential Gauge Error (first author)
    • Separate or Unified Fine-Tuning? A Systematic Comparison of Training Strategies for Multi-Document Domain Applications (second author)

  • Jan. 2026 KSEA SWRC · CSU Fullerton

    • Subliminal Learning in Large Language Models
    • A Design Science Research Approach for Geospatial AI: Predicting Drug Overdose Risk Using Social Media

  • Aug. 2025 AMCIS 2025 · Montréal, Canada

    • A DSR Approach for Geospatial AI: Prediction of Drug Overdoses (paper presentation)

Projects

AI Degree Planning Companion screenshot

AI Degree Planning Companion

An LLM agent that helps CBU students plan courses, supported by the Office of Student Success. Local Mistral via Ollama, LangChain tools over a Neo4j prerequisite graph, PostgreSQL, and FAISS search, streamed to a React chat UI.

LangChainNeo4jFastAPIOllama

Read the write-up →Code on GitHub →

Local Voice Assistant screenshot

Local Voice Assistant

A hands-free assistant that runs fully offline: Vosk speech-to-text with WebRTC VAD, a local Llama 3.2 model through Ollama, and pyttsx3 speech output, with multi-turn memory.

VoskOllamaPython

Read the write-up →

Conceptual illustration of source documents moving through question-and-answer examples, model fine-tuning, and desktop deployment.

Domain Fine-Tuning Pipeline

The CAIMLL lab’s end-to-end workflow for 3B-parameter models: PDF ingestion, source-verified synthetic Q&A, QLoRA/Unsloth training, structured evaluation, and local GGUF deployment.

QLoRAUnslothGGUFvLLM

Training loss versus training steps from the subliminal-learning student model, falling rapidly and then leveling off.

Subliminal Learning Replication

Controlled teacher–student experiments testing whether hidden preferences transfer through filtered number sequences, with LoRA and system-prompt teachers.

Llama 3.2LoRAEvaluation

Code on GitHub →

Gravitational Lensing Simulations

Object-oriented simulations and interactive visualizations of multi-body gravitational lensing, built for computational physics research.

Looping clip from a recorded run of the 2D ray-tracing model: light rays bending as they meet interleaved gravity / anti-gravity rows. Exploratory model, not a physics validation.

PythonSimulationVisualization

Kepler.gl map of the United States with thousands of geolocated social-media posts as colored dots; four treatment-region states shaded in purple.

GeoAI: Drug Overdose Risk Mapping

Geospatial analysis of suspected drug-overdose risk from social-media data: Kepler.gl maps of geolocated posts across the U.S. turned model output into interpretable regional insights. Coauthored the AMCIS 2025 paper and presented the poster at KSEA SWRC.

Fig. 1 from the KSEA SWRC research poster.

Kepler.glGISDeep LearningResearch Poster

Read the paper →

Storm Timeline Analysis (Hydrology) screenshot

Storm Timeline Analysis (Hydrology)

A 28-year daily precipitation record (1996–2024) read timestamp by timestamp: every rainfall spike plotted day by day, with the largest total and peak storms highlighted as validation windows for streamflow models.

PythonTime SeriesValidation

Living Kanto

A persistent Pokémon world where 100 residents, each driven by a local language model, make their own decisions — travelling, working, talking, and pursuing ambitions in a world with consistent rules and lasting consequences.

Replay of recorded simulation activity. Pokémon graphics belong to their owners.

Local LLMsAgentsSimulation

Code on GitHub →

Conceptual illustration of three students' hands holding phones around shared study books, connected by a circular line.

CampusMate

A mobile app that helps college students connect, form study groups, and communicate, with real-time messaging, file sharing, and a group calendar. Started as a CSC322 class project.

FlutterFirebaseMobile

Code on GitHub →

Moments

  • Kip standing with a colleague in front of the AMCIS 2025 Montréal backdrop
    Aug. 2025 AMCIS 2025 in Montréal, presenting our GeoAI paper
  • Kip in a suit standing beside his research poster
    Jan. 2026 Two research posters at KSEA SWRC, CSU Fullerton
  • Cactus × Google DeepMind hackathon title card with a green wireframe hand reaching toward a human hand
    Feb. 2026 3rd place (online division), Cactus × Google DeepMind FunctionGemma Hackathon
  • SO-101 robot arm on a desk next to a monitor showing its simulation
    2026 Physical AI capstone: SO-101 arm assembly and teleoperation setup
  • NVIDIA Jetson Orin Nano developer board on a desk
    Winter break A fully local AI assistant on an NVIDIA Jetson Orin Nano Super
  • Partly assembled Reachy Mini robot on a table
    2026 Assembled Reachy Mini and ran its demos for visiting students
  • Kip and teammates outside the ICPC contest venue
    Nov. 2024 SoCal Regional ICPC with the CBU team

Experience & leadership

  1. Jun. – Sept. 2026

    AI Evaluation & Optimization Intern

    Portfolio Elevation, LLC · Redlands, CA

    • Built a three-part evaluation program: black-box response benchmarks, blind human photo review, and controlled vision-pipeline experiments.
    • Used checked answer keys and repeated runs to catch incomplete responses, negation errors, truncation, and image-orientation failures.

  2. Jan. 2025 – Present

    AI Engineering Intern

    CBU Office of Student Success / Career Center · Riverside, CA

    • Developed an LLM email-drafting system piloted with 3–4 advisors, and a Whisper transcription prototype with projected savings of $6K+ per year.
    • Built a Neo4j/LangChain academic-planning prototype and authenticated mock-interview scheduling apps (Power Apps, Power Automate, SharePoint).

  3. Mar. 2024 – Apr. 2026

    President, CBU ACM

    Founded the Software Engineering Team and mentored 20+ student developers using Scrum and GitHub.

  4. Aug. 2024 – Present

    Software Engineering Tutor, CBU

    Support ~90 students in Java, data structures, debugging, and object-oriented design.

  5. CBU nominee

    Alpha Chi Gaston/Nolle Scholarship

    Selected by CBU for national scholarship consideration.

Toolbox

Languages

PythonJavaC++JavaScript

ML frameworks

PyTorchJAX / HaikuTransformerLensUnslothLangChain

Methods

LoRA / QLoRASparse autoencodersModel evaluationSynthetic data

Inference & systems

OllamavLLMGGUFFastAPINeo4jPostgreSQLDocker

Blog

The AI Degree Planning Companion: Part 1

From idea to architecture: laying the foundation for an LLM-powered degree planning agent at CBU.

LLMsAgentsEducation

Giving My LLM a Voice — Building a Voice-Controlled AI Assistant

Building a fully local, hands-free voice assistant with Vosk, Ollama, and pyttsx3.

LLMsVoiceLocal AI

All posts →

Contact

Let’s talk.

Questions about a project, research collaborations, or opportunities: send me a message. I’ll do my best to get back to you soon.

nankipung@gmail.comLinkedIn