For 5+ years I have built production systems across robotics, fintech, and healthcare. Applying to US CS master's programs to research the reliability of cyber-physical systems, with a PhD as the goal.
Portfolio · hurjun96@gmail.com · LinkedIn · KCC 2024 paper (PDF, in Korean; English summary on pp. 1–2)
Start here: mapf-fleet (simulation-based validation: zero collisions across 66 seeded runs) and mlops (an edge-to-cloud safety-event pipeline that keeps working when the link drops; 33 tests).
Three questions I first met as production problems:
- How do we localize faults across system layers? In the fleets I worked on, one robot command crossed the frontend, backend, MQTT broker, robot ROS stack, and site network. Faults ignored those boundaries.
- How far do early-warning signals in telemetry generalize? At DOGU Robotics, anomalies in MQTT message frequency preceded customer failure reports. I built monitoring on that signal with no new instrumentation.
- What does a passing simulation guarantee about field reliability? At GoLe Robotics I validated a full multi-floor robot delivery in NVIDIA Isaac Sim before on-site runs, and I wrote a multi-robot path-planning simulator from scratch.
Every number below is reported in the linked README (the MediVoice figures come from the paper).
| Repository | What it is | Result |
|---|---|---|
| mapf-fleet | Multi-robot path planning (windowed cooperative A*, Conflict-Based Search) inside a 3D multi-floor site simulator, TypeScript | 0 collisions across 66 seeded runs / 21,000+ ticks; CBS matches throughput at up to ~9× the compute |
| mlops — PPE Watchman | Edge-to-cloud safety-event pipeline: edge detector (stock YOLOv8n, PPE model pluggable) → FastAPI hub → WebSocket dashboard with backpressure | 33 tests over the rules and the ingest → persist → fan-out path; detection accuracy not evaluated |
| attention_is_all_you_need | The Transformer rebuilt from nn.Linear-level primitives, with attention-map plots |
98.2% exact match on a reversal task (~90 s, CPU); in a matched 1,200-step ablation, removing positional encoding drops 91.2% → 0% |
| xLSTM | sLSTM and mLSTM cells from the paper (Beck et al., NeurIPS 2024) | 100% on 64-step recall; a vanilla LSTM stays at chance (6.8% vs 6.25% expected) |
| speculative-decoding-lab | Speculative decoding with distribution-preserving acceptance, checked token-for-token against greedy decoding | 1.80× lossless on Apple M2 MPS (GPT-2-large / DistilGPT-2); 0.70× on CPU (GPT-2-medium / DistilGPT-2); analytic model within 0.35 token/round of measurement |
| stock-forecast-benchmark | 9 forecasting models behind one interface, with data-leakage tests | Prophet ranks first (RMSE 160) on 1962–1992 → 1993–2000 held-out data; explains why the deep models lose here |
| stt-nursing-system — MediVoice | Voice-first nursing documentation; implementation of the KCC 2024 paper | Documentation time −70% on first use and −96% on repeated use (simulated records, 5 trials) |
More: open3d (point-cloud pipeline: RANSAC, DBSCAN, ICP, Poisson) · opencv (Faster R-CNN detection with a tested ROI rule engine) · setpoint (RL from scratch: value iteration, REINFORCE) · futurescole (seeded event pipeline on PostgreSQL) · gre (adaptive testing engine with convergence simulation)
- Xpanner — Senior Software Engineer · 2026.08 – Present. Control and management software for automation retrofits on heavy construction equipment: machine signals to operator consoles and the cloud (AWS IoT Core, ECS, Lambda, RDS). Simulation-based validation workflows in Isaac Sim / Isaac Lab.
- GoLe Robotics — Full-Stack & Robotics Simulation Engineer · 2026.05 – 2026.08. Fleet management system for delivery robots (VDA5050 over MQTT, WebSocket, PostgreSQL). Built an Isaac Sim environment where the robot's real ROS2 nodes run unmodified; validated a full multi-floor delivery before on-site runs.
- Niflers (fintech) — Full-Stack Engineer, sole developer · 2024.01 – 2025.10. Shipped CRISK, Credit Imbalance Tracker, and Credivalue. 200,000+ auto-generated carbon-risk reports; report generation cut from 10 min to 2; daily pipelines on AWS ECS + Step Functions.
- DOGU Robotics — Web Development Lead, robot fleet control · 2021.11 – 2023.05. 30+ robots across 10+ sites for Hyundai, SK hynix, and GS EPS. Debugged incidents across frontend, backend, MQTT broker, and robot ROS layers; turned MQTT message-frequency anomalies into proactive monitoring.
- Earlier (2021): Mobile Entropy — ERP web services for Incheon City Gas · NSG — C++ digital twin of nuclear-plant thermal-hydraulic logic, with K-water.
- B.S. in Computer Science & Engineering, Chungnam National University, 2016.03 – 2025.02. Completed while working full-time as a software engineer from 2021.
- H. Choi, Y. Kim, J. Hur, J. Park. Research on Work Efficiency Using AI Speaker-based Nursing Record Assistance System. Proc. KCC 2024 (KIISE), pp. 2015–2017. Third of four authors: system implementation and evaluation design. Paper in Korean. PDF with English summary (pp. 1–2) · code
- Grand Prize, KIISE Software Implementation Competition, 2017 — Reacord, real-time lecture transcription for deaf and hard-of-hearing classmates.
Python · TypeScript · C++ · Java · PyTorch · ROS2 · NVIDIA Isaac Sim / Isaac Lab · MQTT / VDA5050 · WebRTC · FastAPI · Spring Boot · React · PostgreSQL · MySQL · AWS · Docker