I study AI Computing at KAIST (B.S. 2026–2030, with an intended double major in Business & Technology Management) and do undergraduate research in the Interactive Computing Lab with Prof. Uichin Lee. I build machine-learning systems end to end, from data collection to a working product, and I care most about the evaluation: pre-registered comparisons, leakage checks, baselines and confidence intervals, with the limits stated next to the numbers.

LinkedIn · marlen.melis@kaist.ac.kr · Daejeon, South Korea

Rescue Router · live viewer
Flood maps from drone photos and truck → boat → on-foot rescue routes to every house, learned from 50 labelled photos. A pre-registered comparison found that ordinary fine-tuning beats the course's few-shot generalist from about 20 labels: mIoU .617 vs .507 on all 448 FloodNet test photos, with houses wrongly shown as truck-reachable falling from 11.1% to 6.8%. A 13 MB model maps your own photo in the browser.
_{PyTorch · torchvision · ONNX Runtime Web · Leaflet · JavaScript — KAIST AIC.20000, Fall 2026}

QuantPulse · live demo
A $0 equity-research and portfolio-construction engine for the S&P 500: eight free data sources, seven-category composite scoring, walk-forward backtests with bootstrap confidence intervals, forecasts graded against a random walk, and three portfolio optimisers (mean-variance, HRP, Black-Litterman). 21,000+ lines of engine code and 2,000+ tests, including property-based ones; a scheduled pipeline refreshes the demo every weekday.
_{Python · pandas · scikit-learn · statsmodels · FastAPI · React/TypeScript · SQLite — personal project, 2026}

Code Review AI
Can an LLM review pull requests better than a model trained on the project's history? I mined 1,494 PRs from five large open-source repositories, a quarter of them written by an AI agent, and ran four experiments. The best LLM configuration matched, not beat, a Random Forest (93.75% accuracy, identical confusion matrix), and only the self-reflection and multi-turn prompts ever caught a bad AI-written PR. Ships as a FastAPI backend and a VS Code extension.
_{Python · scikit-learn · LLM prompting · FastAPI · TypeScript — HIT Global Summer School, July 2026}

Theme Worlds · live viewer
Seamless 360° panoramas from stock Stable Diffusion v1.5 with a MultiDiffusion sampler written from scratch: circular windows make the wrap-around a property of the sampling, and multi-branch conditioning avoids CLIP's 77-token truncation. Mean seam ratio 1.12 over 28 panoramas (about 1 means the join looks like any other pair of columns); in paired controls, switching circular windows off raises it from 1.05 to 7.94. Reruns are byte-identical.
_{PyTorch · diffusers · CLIP — KAIST AIC.20000, Fall 2026}

ModuleMate
A local AI study workspace for your own notes, slides and scans: answers with inspectable sources, learning cards bound to exact source spans, and a planner that turns dates in the material into editable calendar events. Runs a quantised 9B model on Apple Silicon with local OCR, and adds its own trained relevance, answerability and ranking heads; 480 tests.
_{Python · MLX · sentence-transformers · SQLite FTS5 · Streamlit — began as a HIT Summer School project, course grade 97/100}

• Undergraduate researcher, Interactive Computing Lab, KAIST (Sep 2026 – present)
• Sole-author journal article: "How Did Personal Loans Influence the Well-Being of Financially Unstable Families in Kazakhstan Over the Last Five Years?", Central Asian Scientific Journal 4(23), pp. 12–21, 2024 (in Kazakh). A 101-respondent survey and two semi-structured interviews, designed, run and analysed alone.

• KAIST, B.S. in AI Computing, 2026 – present
• Harbin Institute of Technology, Global Summer School, July 2026: AI & Machine Learning (97/100), Basics of Machine Learning (91/100), Intelligent Software Engineering practice (90/100)

Languages: Python, TypeScript, JavaScript, SQL, C/C++
ML: PyTorch, diffusers, scikit-learn, statsmodels, sentence-transformers, FAISS, MLX, ONNX Runtime
Systems: FastAPI, SQLAlchemy/Alembic, React, Streamlit, Gradio, SQLite, GitHub Actions
Practice: pytest, Hypothesis, Playwright, ruff, mypy, pre-registration, bootstrap confidence intervals
Spoken: English (IELTS 8.0), Kazakh and Russian (native)