Gio Kim

Research · Engineering · Competition

Projects

Code you can run. Everything here has a repository.

BioGemma — Gemma 4 E2B Biology LoRA Experiment

Fine-tuned LoRA adapters on Gemma 4 E2B for biology Q&A, then documented why the benchmark numbers didn't move — an honest negative-result writeup.

pytorch · lora · gemma · fine-tuning · biology
Read the full write-up

Fine-tuned rank-8 LoRA adapters on about 10,000 biology/biomedical examples for 500 steps on an NVIDIA T4.

Held-out loss dropped from 9.55 to 7.35 (23%), but overall benchmark accuracy stayed flat at 56.5% — documented the negative result and the evaluation mismatch between training loss and benchmark performance rather than overstating the win.

WebLLM — Browser-Based Vision-Language Model Chrome Extension

A Chrome extension that runs a vision-language model (Gemma) entirely client-side via WebGPU/ONNX Runtime — on-screen visual Q&A and chat with no backend server.

chrome extension · webgpu · onnx runtime · gemma · transformers.js · manifest v3

Protein Sequence Transformer — Evo-Biased Generator

An autoregressive PyTorch Transformer that generates LuxR-family protein sequences, with a learned pairwise position bias trained on UniProt data.

pytorch · transformer · protein-design · uniprot · bioinformatics
Read the full write-up

Developed a Transformer-based autoregressive language model in PyTorch to generate protein-like amino acid sequences for the LuxR regulatory protein family, incorporating a learned pairwise interaction bias between sequence positions.

Trained on bacterial regulatory protein sequences retrieved from UniProt (with fallback datasets) and built a heuristic ranking system for generated sequence outputs — without claiming structural validity.

Simplified Kinetic Modeling of the RhlI/RhlR Quorum Sensing System

A reaction-based systems biology model of the RhlR/RhlI quorum sensing pathway in Pseudomonas aeruginosa, built with Tellurium and Hill-function kinetics.

tellurium · systems-biology · kinetic-modeling · quorum-sensing
Read the full write-up

Developed a reaction-based systems biology model of the RhlR/RhlI quorum sensing pathway using Tellurium and Hill-function kinetics.

Investigated positive-feedback-driven signal amplification through simulations of C4-HSL production, receptor binding, and degradation dynamics; published the full computational analysis and findings on Cells and Civics.

Structural Prediction of Babesia bovis HAP2 Protein

Predicted the full-length Babesia bovis HAP2 structure with AlphaFold2 and validated it against published ectodomain data.

alphafold2 · chimerax · structural-biology · protein-structure
Read the full write-up

Predicted the full-length Babesia bovis HAP2 structure via AlphaFold2 and aligned it against published ectodomain data (Rahman et al.), obtaining a 1.263 Å RMSD.

Analyzed conservation of membrane-fusion motifs despite about 27% sequence identity, and published the methods, alignment, and interpretation on Cells and Civics.

Structural Prediction of LasR LBD–QslA Complex

Predicted the 3D structure of the LasR ligand binding domain–QslA complex in P. aeruginosa using AlphaFold2 and analyzed dimerization and ligand-binding features in ChimeraX.

alphafold2 · chimerax · structural-biology · quorum-sensing
Read the full write-up

Predicted the 3D structure of the LasR ligand binding domain–QslA complex in Pseudomonas aeruginosa using AlphaFold2.

Analyzed structural features relevant to dimerization and ligand binding in quorum sensing regulation using ChimeraX; published on Cells and Civics.

Atlas — Local-First Personal AI Assistant

A local-first personal AI assistant: 63 FastAPI routes, 95 offline tests, persistent Obsidian-vault memory, and voice/chat/WhatsApp interaction.

python · fastapi · mcp · local-first · agents
Read the full write-up

Shipped 3 public versions in 8 months (Dec 2025 - Jul 2026). V3 exposes 63 FastAPI routes and includes 95 offline tests covering memory, routing, permission gates, workflows, and OS safety.

Built persistent Obsidian-vault memory, voice/chat/WhatsApp interaction, multimodal attachments, and audited tool use across local and API-hosted models. Evolved from the Jarvis automation prototype (Nov-Dec 2025).

Designed privacy- and cost-conscious on-device workflows with local wake-word/STT, plain-Markdown memory, hybrid BM25/vector RAG, MCP tools, and provider failover, allowing supported tasks to avoid cloud data transfer and per-request API costs.

Cells and Civics + Freelance Web Development

Founded Cells and Civics, a research platform publishing computational biology studies and AlphaFold structural analyses, and designed and developed full websites for 4+ nonprofit organizations.

web-development · freelance · nonprofits · publishing
Read the full write-up

Delivered end-to-end websites for 4+ nonprofits, independently owning information architecture, interface/content design, implementation, deployment, and launch.

Founded Cells and Civics to publish independent computational biology analyses, including AlphaFold structural studies and systems-biology modeling.

Breast Cancer CNN Classifier

A benign/malignant breast cancer image classifier built in PyTorch, with an augmentation pipeline, CrossEntropyLoss, and the Adam optimizer.

pytorch · cnn · computer-vision · medical-imaging
Read the full write-up

Built a benign/malignant breast cancer image classifier in PyTorch, with a data augmentation pipeline, CrossEntropyLoss, and the Adam optimizer.