Freshman Software Engineering Student @ Ajou University
Solo-building ML/physics-simulation projects outside coursework — in short, intense bursts, not a steady grind.
1st-year Software Engineering student at Ajou University (started March 2026). Every project below is solo, unreleased, and unstarred — built to learn something specific, not to ship a product. Development happens in bursts: a few weeks of daily commits on one repo, then months of nothing, then a burst on the next one.
- 🔬 What I've actually built so far: gradient-diagnostics tooling, PDE solvers (FNO/PINN), classical-ML-vs-GNN experiments, a chest X-ray classifier prototype, and a from-scratch OS scheduler/memory simulator
- 🏥 Where I'm aiming, not where I am: medical imaging, computer vision, and data science. Starting the Medical AI micro-degree this coming semester — the draw is access to Ajou University Hospital's clinical infrastructure, which is the part that's hard to replicate elsewhere. Nothing shipped in that direction yet
- 📐 How I actually work: derive or read the paper before importing the library. Most performance numbers in the projects below are self-measured inside each repo, not independently verified — flagged explicitly there rather than left to imply more than they show
| Project | What it is | Stack |
|---|---|---|
| Gradient Pathology | PyTorch library that profiles gradient statistics during training to catch vanishing/exploding gradients and dead neurons — layer-wise heatmaps, Sankey diagrams, a live Streamlit dashboard, and a 7-rule diagnostic engine with code-fix suggestions. Has tests and an in-repo benchmark (under ~3% training overhead). Zero stars, zero outside users — a solo tool, not an adopted library. | PyTorch Plotly Streamlit Python |
| Physics-Informed ML | Fourier Neural Operator and PINN implementations for the heat equation, Burgers', and Navier-Stokes, wrapped in a FastAPI backend and a React/Three.js 3D frontend, plus uncertainty-quantification experiments (Bayesian PINN, deep ensembles). The "100–1000× faster than solvers" number compares a trained model's inference time to a full classical solve — a self-reported, in-repo comparison, not a third-party benchmark. | PyTorch FastAPI React Three.js Terraform |
| ML Gradient Descent Viz | Started as a from-scratch NumPy backprop/MLP implementation, grew into a small optimizer library (SGD through AdamW) with convergence-theory notes and JAX/CuPy GPU benchmarks. The "deep non-convex" phase mentioned in the repo is still unbuilt — everything working today is convex/near-convex. | NumPy JAX Python |
| Physics-Informed Optimizer | JAX/Flax framework for solving PDEs with physics-informed neural nets (heat equation, Navier-Stokes) — PINN-specific training tricks (curriculum learning, adaptive loss balancing), multi-GPU training, a Streamlit dashboard. Despite the name, it's not a general-purpose optimizer that uses physical priors; it's a PDE solver. | Python JAX Flax |
| Chemical Reaction Rate Prediction | Started as a high-school chemistry class project. Now compares RandomForest/XGBoost against GNNs (GCN, GAT, GIN, MPNN) for predicting reaction rates from molecular structure, with a FastAPI + React interface. Headline numbers in the repo (R² 0.985, +18% from a hybrid model) are self-reported from commit messages, not from a published or independently reproduced result. | Scikit-learn PyTorch (GNN) FastAPI React |
| Project | What it is | Stack |
|---|---|---|
| Ajou Dorm Finder | Unofficial dorm-eligibility checker and assignment-score calculator for Ajou University students, with per-semester facility data, room-type charts, and a countdown to the new dorm building. No backend — all data is hardcoded per semester and updated by hand. | React TypeScript Vite |
| AI Disease Classifier | Chest X-ray pneumonia-classification prototype — Flask API, ONNX Runtime inference, Grad-CAM heatmaps for explainability. The ONNX step is a format conversion, not quantization; there's no training code in the repo to back a class-imbalance fix. The project's own README already flags it as an educational prototype with no FDA/CE clearance. | Flask ONNX JavaScript |
| Mini OS Simulator | CLI simulator of process scheduling (FCFS, SJF, Round Robin, Priority), paging and page replacement (FIFO/LRU), and syscalls (SLEEP/IO/FORK/EXIT) — built to learn OS fundamentals by implementing them, not to run anything real. | Python C |
| NPU Simulator | Repo created, nothing committed yet. Listed as an open placeholder for a hardware-inference-acceleration idea I haven't started, not as finished work. | — |
| Project | What it is | Stack |
|---|---|---|
| AI Cyberbullying NLP Analysis | One-day project: TF-IDF + scikit-learn classifier for Korean toxic-comment detection, trained on 80k rows sampled from an AI Hub ethics dataset. No transformer model — that's future work noted in the README, not something built. | Python Scikit-learn |
| Category | Technologies |
|---|---|
| Languages |     |
| ML & Science |     |
| Backend & APIs |    |
| Frontend |    |
| DevOps & Cloud |     |
| Period | Institution | Details |
|---|---|---|
| 2023 – 2025 | Eunhye High School | — |
| 2026 – | Ajou University | Dept. of Software — Software & Computer Engineering major |
| 2026 – | Ajou University | Medical AI micro-degree — coursework starts this semester, in progress |
Learning log at sinsangwoo/TIL — notes on algorithms, ML theory, and tools, capped at 5 lines a day. Automated and daily since mid-August 2026; before that there's a 5-month gap where nothing was written.
16 public repos, 0 stars — each built to learn one thing, not to ship. That part's next.