+------------------------------------------------------------+
| KAPASIQUE :: ARTEM SVINOBOEV |
| ML / COMPUTER VISION / FULL-STACK / AGENTIC ENGINEERING |
+------------------------------------------------------------+
I turn ambitious ideas into systems that work.
architecture first / agents with supervision / proof over hype
I'm a computer engineering student from Yakutsk, currently studying between NEFU and Jiamusi University in China. My main lane is machine learning and computer vision; my unfair advantage is being able to carry the same idea all the way from an experiment to a usable product.
I have spent roughly four years building for the web, then moved deeper into ML, CV, mobile, and agent-driven development. The long game is graduate study in AI / CS in southern China — and a career built around hard technical problems, not one narrow framework.
Yakutsk ---> China ---> Shenzhen / Guangzhou
web ML/CV research + products
I work end to end: turn an ambiguous problem into system boundaries, build the critical path, design the evidence that can prove it works, and trace failures across data, models, APIs, and runtime behavior. I keep reviewing until I can explain the trade-offs, reproduce the result, and operate what ships.
FRAME -> requirements / constraints / failure modes
DESIGN -> architecture / data contracts / validation
BUILD -> critical paths / integrations / interfaces
PROVE -> tests / CV / ablations / instrumentation
DEBUG -> data / model / API / runtime
SHIP -> deploy / monitor / document / own
AI tools accelerate exploration and implementation. They do not own the acceptance decision: I verify the code, the measurements, and the production behavior before I put my name on the result.
| Project | What happened |
|---|---|
stellar-class-prediction-s6e6 |
LGBM + XGB + CatBoost + RealMLP stacker for stellar classification. 0.9711 balanced accuracy, top ~8%. |
f1-pitstop-prediction-s6e5 |
OOF-validated GBDT / RealMLP blend. 0.9545 private AUC, top ~7%. |
SMILES-2026 |
Zero-order ResNet18 fine-tuning on CIFAR-100: 49.68% top-1 with no gradient computation. |
kaggle-dominator |
An open Claude Code skill for disciplined, evidence-driven Kaggle work. |
trustlens |
Multi-agent BI system that verifies numbers before presenting them. Built with Google ADK + MCP. |
second-look-triage |
Clinically grounded ER triage safety net with calibration, red-flag NLP, and fairness auditing. |
PaperCV |
Real-time attention and gaze monitoring with MediaPipe, OpenCV, FastAPI, and React. |
yakutsk-city |
Production website for the Yakutsk city IT department, built with Next.js and React. |
ML / CV Python PyTorch scikit-learn LightGBM XGBoost OpenCV
PRODUCT TypeScript React Next.js React Native Node.js FastAPI
SYSTEMS Go Docker Vercel Linux GitHub Actions
AGENTIC Claude Code Codex MCP Google ADK DeepSeek API
Tools change. The job stays the same: choose the right abstraction, find the failure modes, and get the thing across the finish line.
[ learning ] Mandarin + preparation for graduate admissions
[ building ] ML/CV experiments, AI tools, and full-stack products
[ refining ] agent orchestration without surrendering engineering judgment
[ offline ] running toward a half-marathon / lifting / Mercedes enthusiast
If you're working on an ML system, computer-vision product, or a serious agent-driven workflow, take a look through the repositories or reach me here on GitHub.
BUILD REAL SYSTEMS :: KEEP RECEIPTS :: STAY CURIOUS