I build software that keeps robot fleets running in the real world. Over 5+ years I have shipped production systems across robotics, fintech, and healthcare — robot fleet-control platforms, financial data pipelines, and clinical software. I currently build the fleet management system for autonomous mobile robots at GoLe Robotics and validate robot behavior in NVIDIA Isaac Sim before deployment.
I am applying to US CS master's programs to study the reliability of cyber-physical systems — fault diagnosis, failure prediction, and simulation-based validation — with the goal of continuing on to a PhD.
→ Portfolio with demos and screenshots · Undergraduate thesis (PDF)
Three questions, each of which I first met as an operational problem in production:
- Fault diagnosis across system layers. In the fleets I operated, a single command crossed the control frontend, backend, MQTT broker, robot ROS stack, and site network — and failures rarely respected those boundaries. I am interested in observability and fault-localization techniques that trace causes across layers.
- Failure prediction from operational telemetry. At DOGU Robotics I found that anomalies in MQTT message frequency repeatedly preceded customer failure reports, and built monitoring around that signal with no added instrumentation. I want to understand how far such early-warning indicators generalize across systems and failure types.
- Simulation-based validation before deployment. I validate robot behavior in Isaac Sim and am extending that work to Isaac Lab with domain randomization; I also wrote a multi-robot simulator (mapf-fleet) for pre-deployment fleet sizing. I am interested in scenario generation and in quantifying what simulation results guarantee about field reliability.
If you have five minutes: read mapf-fleet (a from-scratch multi-robot path-planning engine, benchmarked) and speculative-decoding-lab (LLM inference acceleration, measured against its analytic model).
Every number below is recorded in the linked repository's README, and each repository documents how to reproduce or verify it.
| Repository | What it is | Evidence |
|---|---|---|
| mapf-fleet | Real-time 3D simulation of multi-robot fleets on multi-floor construction sites. A MAPF engine written from scratch in TypeScript — windowed cooperative A* (WHCA*) over a space-time reservation table, plus optimal Conflict-Based Search — with capacity-limited elevators and an analytical fleet-size optimizer validated against measured throughput. | Zero collisions across 66 seed-swept headless runs / 21,000+ simulation ticks; CBS matches prioritized planning's throughput at up to ~9× the compute. |
| mlops · PPE Watchman | Edge-to-cloud construction-site safety monitoring: YOLOv8 inference on edge clients that stay autonomous offline, lightweight violation events to a FastAPI hub, and live WebSocket fan-out to a Next.js dashboard with bounded-queue backpressure. | 33 passing tests covering the violation-rule logic and the API hub — event ingestion → persistence → WebSocket fan-out with backpressure. |
| Repository | What it is | Evidence |
|---|---|---|
| attention_is_all_you_need | The original Transformer built from nn.Linear-level primitives — multi-head attention, masking, the Noam schedule, label smoothing — with cached attention-map visualization. |
98.2% exact sequence match on a reversal task, reproducible in ~90 s on CPU; a controlled ablation (identical seeds and initialization) removing positional encoding collapses exact match from 91.2% to 0% in the ablation configuration. |
| xLSTM | sLSTM and mLSTM cells with exponential gating and stabilized state (Beck et al., NeurIPS 2024); the parallel mLSTM form is unit-tested to match the O(L) recurrence to within 1e-6. | 100% accuracy on 64-step recall, where a vanilla LSTM never beats chance (~6.25%) even with up to 2× the parameters. |
| speculative-decoding-lab | Speculative decoding with distribution-preserving acceptance sampling, verified bit-for-bit against greedy decoding, plus draft-length and temperature sweeps compared against the analytic acceptance model. | Up to 1.80× lossless speedup on Apple Silicon MPS and 0.70× on CPU — consistent with the memory-bandwidth-bound regime analysis; the analytic tokens-per-round model matches toy-backend measurements within 0.35 token. |
| stock-forecast-benchmark | Config-driven benchmark of 9 forecasting models (ARIMA, ETS, Prophet, XGBoost, LightGBM, LSTM, GRU, TCN, Transformer) behind one interface, with data-leakage safety enforced by tests and a documented limitations section. | Prophet ranks first of the 9 models across MAE, RMSE, and MAPE on 1962–1992 → 1993–2000 held-out data, with an analysis of why the deep models underperform here. |
| Repository | What it is | Evidence |
|---|---|---|
| stt-nursing-system · MediVoice | Voice-first nursing documentation: an AI speaker asks assessment questions (TTS), transcribes the answers (STT), and normalizes them into structured records. The working implementation of my undergraduate thesis. | In the thesis evaluation, the assisted workflow cut documentation time for a repeated assessment by up to 96%. |
More: open3d — the classic point-cloud/LiDAR pipeline (RANSAC, DBSCAN, ICP, Poisson meshing) with a headless test suite · opencv — Faster R-CNN person detection with a testable ROI intrusion-rule engine · setpoint — RL fundamentals (value iteration unit-tested against hand-derived optima, plus REINFORCE) · futurescole — deterministic, seeded event pipeline on PostgreSQL + Docker with hermetic tests · gre — adaptive testing engine evaluated by convergence simulation against a synthetic learner.
Fleet management system (FMS) for autonomous mobile robots on construction sites (golerobotics.com).
- Real-time control and monitoring over the MQTT-based VDA5050 standard with WebSocket live feeds; PostgreSQL backend with role-based access control and centralized logging that distinguishes network, authentication, and database failures.
- Built an Isaac Sim environment where the company's AMR (CAD → URDF) runs its real ROS2 nodes unmodified with Nav2 — validated a complete multi-floor delivery sequence (elevator boarding, floor-map switch) end to end before on-site runs. Now extending to Isaac Lab reinforcement-learning pipelines with domain randomization.
Sole owner of frontend, backend, infrastructure, and deployment. Shipped CRISK, the Credit Imbalance Tracker, and Credivalue.
- Auto-generated 200,000+ corporate carbon-risk reports; cut report generation from 10 minutes to 2 by refactoring a 20,000+ line monolith; daily financial-data pipelines on AWS ECS + Step Functions.
- Diagnosed and resolved two production incidents — a data-integrity failure during 20-process parallel loading, and a framework upgrade that broke every API — experience that shapes how I design for data integrity and observability.
Fleet-control platform for 30+ robots across 10+ sites (hospitals, factories, outdoor) for clients including Hyundai, SK hynix, and GS EPS.
- Debugged incidents end to end across the control frontend, backend, MQTT broker, and robot ROS layers; turned the MQTT message-frequency anomaly described under Research Interests into proactive monitoring, so engineers responded before customers reported failures.
- Shipped branching diagnostic scripts that field engineers used to localize faults by elimination (battery → network → software); WebRTC video streaming, Grafana dashboards, and a Vue.js → React migration standardized across 10+ site deployments.
Earlier: Mobile Entropy — ERP web services for Incheon City Gas (2021) · NSG Co., Ltd. — C++ digital twin of nuclear-plant thermal-hydraulic calculation logic, in collaboration with K-water (2021).
- B.S. in Computer Science & Engineering, Chungnam National University · 2016.03 – 2025.02 — completed while working full-time as a software engineer from 2021 onward. Thesis: an AI-speaker nursing-documentation assistant, evaluated in a time-and-motion study — thesis PDF · implementation.
- Grand Prize, KIISE Software Implementation Competition (2017) — Korean Institute of Information Scientists and Engineers, for Reacord, a real-time lecture-transcription app built and released for deaf and hard-of-hearing classmates.
Languages
ML & Robotics
Real-time / Systems
Backend & Frontend
Cloud, Infra & Data
