B.S. Artificial Intelligence Computing + Business & Technology Management — KAIST, Daejeon, South Korea (2026–2030)
Nazarbayev Intellectual School of Physics & Mathematics, Almaty · HIT Global Summer School, Harbin
| Project | What it is | Stack |
|---|---|---|
| quantpulse · live demo | Equity research and portfolio-construction engine. Eight free data sources into a 23-table schema, seven-category composite scoring, walk-forward backtesting, and three portfolio optimisers (mean-variance, Hierarchical Risk Parity, Black-Litterman). ~17k lines of engine code, 1,466 automated tests including property-based. | Python · scikit-learn · statsmodels · FastAPI · React/TS · SQLite |
| code-review-ai | Four experiments on whether an LLM can predict and review pull requests better than a trained model: a 1,494-PR dataset I mined from five large OSS repositories, merge-prediction models, an LLM review pipeline, and a FastAPI backend plus a VS Code extension that wrap the best of both. | Python · scikit-learn · FastAPI · TypeScript |
| modulemate | Retrieval-grounded course assistant that answers only from staff-issued material and cites the slide it came from. Hybrid dense + BM25 retrieval, cross-encoder re-ranking, and three independent refusal layers. | Python · FAISS · sentence-transformers · Streamlit |
| theme-worlds · live viewer | Seamless 360° panoramas from stock Stable Diffusion v1.5. A MultiDiffusion sampler written from scratch, with circular windows so the wrap-around loop is a property of the sampling rather than a post-process, and multi-branch conditioning so two themes share one world without CLIP's 77-token limit truncating either prompt. No training, zero source images. | Python · PyTorch · diffusers · CLIP |
| rescue-router · live viewer | Flood maps from drone photos and rescue routes to every house: a five-class segmentation learned from 50 labelled photos, and a router that sends a truck on dry ground, a boat across the water and a team on foot for the last stretch, with a confidence tier on every truck route. A pre-registered comparison of a few-shot generalist (Chameleon) with ordinary fine-tuning on the same photos decided which model ships; a 13 MB model maps your own photo in the browser. | Python · PyTorch · torchvision · ONNX Runtime Web · JavaScript · Leaflet |
- quantpulse — Sharpe and CAGR are reported as moving-block bootstrap confidence intervals, never as a bare point estimate. Every forecast (ARIMA/SARIMA, gradient boosting, Monte Carlo) is graded out-of-sample against a random-walk baseline before it is allowed to ship.
- code-review-ai — the best LLM configuration matched rather than beat the trained Random Forest: 93.75% accuracy, 0.816 macro-F1, an identical confusion matrix. That was the opposite of my hypothesis, so it became the headline finding.
- modulemate — zero fabricated citations and a 1.000 refusal rate on the out-of-scope set across a 63-question gold set, with the false-refusal rate (0.038) published beside it, because a refusal rate on its own can be maximised by refusing everything.
- theme-worlds — a mean seam ratio of 1.12 over 28 panoramas, where 1.0 means the wrap-around join is indistinguishable from any other pair of columns, against 7.94 for the same seeds with circular windows switched off. The same seeds rendered in two separate Colab sessions produced byte-identical PNGs.
- rescue-router — the CPU ensemble that ships reaches mIoU .617 [.583, .652] on all 448 FloodNet test photos against .507 for the few-shot model it replaced, and shows a house as truck-reachable when it is not 6.8% of the time against 11.1%. The few-shot model still wins with 1–5 labelled photos. The questions, runs and decision rules were committed before the first GPU run, so the outcome could not shape the test.
Sole author, peer-reviewed — "How Did Personal Loans Influence the Well-Being of Financially Unstable Families in Kazakhstan Over the Last Five Years?", Central Asian Scientific Journal, Vol. 4(23), pp. 12–21 (2024, published in Kazakh). A 101-respondent survey and two semi-structured interviews, designed, fielded and analysed alone.
Languages — Python, TypeScript, SQL, C
ML and data — PyTorch, diffusers (Stable Diffusion), scikit-learn, statsmodels, pandas, NumPy, FAISS, sentence-transformers
Systems — FastAPI, SQLAlchemy, Alembic, React, Streamlit, SQLite
Engineering — pytest, Hypothesis (property-based testing), GitHub Actions CI, ruff, mypy, pre-commit
Quantitative — portfolio optimisation (MPT, HRP, Black-Litterman), backtesting, bootstrap confidence intervals, Monte Carlo, ARIMA/SARIMA
Spoken — English (IELTS 8.0), Russian (native), Kazakh (native)