Skip to content
View Kipung's full-sized avatar
12 followers · 6 following

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2

Block or report Kipung

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Kipung/README.md

Ki-Pung (Kip) Park. Fine-tuning, evaluating, and running language models locally.

Website  ·  CV  ·  LinkedIn  ·  Email  ·  AMCIS 2025 paper  ·  Publications

Now: building the CAIMLL lab's fine-tuning pipeline, leading a subliminal-learning study through revision, and studying rainfall estimation for flood prediction.

Figures

Fig. 1. Living Kanto: A persistent world where 100 residents, each run by a local language model, travel, work, talk and pursue their own ambitions. Fig. 2. Domain Fine-Tuning Pipeline: Reference documents become source-verified training data, checked for evidence coverage, answer quality and reproducibility. Fig. 3. Subliminal Learning in LLMs: Can a hidden love of owls pass from teacher to student through number sequences? The teacher picked owls 53% of the time; students barely moved. Fig. 4. GeoAI: Overdose Risk Mapping: A deep model predicts areas of suspected drug-overdose risk from geolocated posts; Kepler.gl maps turn its output into regional insight.

More projects

AI Degree Planning Companion: An agent that plans courses over a Neo4j prerequisite graph, on local models, streamed to a React chat UI. Terminal Browser Split: A live, clickable browser preview in a native terminal split, opened by you with a hotkey or by any coding agent with one command.

Rainfall Transfer Between Gauges: Hourly rainfall from one gauge estimated from another across 28 years of records; held-out event hours reach R² = 0.58. CBU ACM Website: Built by the Software Engineering Team I founded, with Brandon Magana and the SET crew, using Scrum and GitHub flow.

6 more projects

Tracr → TransformerLens: Compile a RASP program with Tracr, convert it to TransformerLens, check both decode alike. GeGLU vs ReLU: Feed-forward activations compared across widths, with bootstrap confidence intervals.

Local Voice Assistant: Speech in, local Llama 3.2 out, spoken back; nothing leaves the machine. BAY-min: A robot math tutor that talks, watches the student’s work, and nods along.

CampusMate: Study groups, real-time messaging, file sharing and a group calendar for students. FunctionGemma Hackathon: 3rd place in the online division of the Cactus × Google DeepMind FunctionGemma Hackathon.

In motion

Living Kanto replay: residents walk through town, each deciding where to go next Claude Code gets a plain-English request, edits the page, then reloads, fills in and clicks it in the terminal-browser split beside it Gravitational lensing simulation: light rays bend as they reach rows of masses
Living Kanto · 100 local-LLM residents Terminal Browser · Claude Code drives a live page Gravitational lensing · physics research

Photos

Conferences, competitions, robots and hardware

Conferences & competitions: AMCIS 2025, Montréal, Presenting our GeoAI paper; KSEA SWRC, CSU Fullerton, Two research posters, Jan. 2026; SoCal Regional ICPC, With the CBU team, Nov. 2024

Robots & hardware: SO-101 arm, Physical AI capstone: assembly and teleoperation; Reachy Mini, Assembled and demoed for visiting students; Jetson Orin Nano Super, A fully local AI assistant on the edge

Timeline

Fig. 5. Timeline of research, work and leadership from 2024 to now.

Publications & talks

  • Corso, A., Asariah, J. S., Park, K., Corso, N. A., & Lee, E. (2025). A DSR Approach for Geospatial AI: Prediction of Drug Overdoses. AMCIS 2025, Montréal. Paper
  • Park, K., Bergquist, G., Lee, P., & Sanders, B. Replicating Subliminal Learning in LLMs via Distillation and LoRA Fine-Tuning. Under review, Aletheia: The Alpha Chi Journal of Undergraduate Scholarship.
  • SCCUR 2026, San Diego State (Nov. 21): Towards Modeling Runoff and Flood Prediction over Unreliable Precipitation Gauges (first author) and Separate or Unified Fine-Tuning? A Systematic Comparison of Training Strategies for Multi-Document Domain Applications (second author).
  • KSEA SWRC 2026, CSU Fullerton: Subliminal Learning in Large Language Models; A Design Science Research Approach for Geospatial AI.
  • AMCIS 2025, Montréal: paper presentation.
  • 🏆 3rd place (online division), Cactus × Google DeepMind FunctionGemma Hackathon.

Toolbox

Python, PyTorch, C++, Java, JavaScript, TypeScript, React, FastAPI, Flutter, Firebase, Docker, PostgreSQL, Linux, Git

ML PyTorch · JAX / Haiku · TransformerLens · Unsloth · LangChain  ·  Methods LoRA / QLoRA · sparse autoencoders · evaluation · synthetic data  ·  Inference Ollama · vLLM · GGUF  ·  Data Neo4j · PostgreSQL · FAISS

Terminal

An animated terminal introducing Ki-Pung Park, research interests, and projects. Build. Test. Understand. Repeat.

Riverside, CA · Interests: LLM architectures, vision-language-action models, physical AI

Pinned Loading

  1. WeHateOwls WeHateOwls Public

    Replicating subliminal learning in LLMs: can an owl preference pass from teacher to student through number sequences? LoRA and system-prompt teachers on Llama 3.2.

    Jupyter Notebook 1

  2. SET_PROJECT_3 SET_PROJECT_3 Public

    Website for the ACM chapter at California Baptist University, built by the ACM Software Engineering Team.

    HTML 1

  3. BAY-min_personal_tutor BAY-min_personal_tutor Public

    BAY-min: a robot math tutor on Reachy Mini, with Gemini Live voice, camera checks of student work and expressive head motion.

    Python

  4. degree_planning_guide degree_planning_guide Public

    California Baptist University - Office of Student Success degree planning guide

    Python

  5. living-kanto living-kanto Public

    Persistent Living Kanto simulation with autonomous local-model residents

    Python

  6. terminal-browser terminal-browser Public

    Live, clickable browser preview in a native terminal split beside your coding agents — Ghostty, kitty, WezTerm, tmux on macOS + Linux. Addon for terminal-browser.dev.

    Shell