Skip to main content

About Me & Global Dev Blog

Welcome to Global Dev Blog, an engineering lab notebook and technical publication authored by Jihwan Park (jbjihwan).

This blog documents hands-on engineering journeys, real-world deployment challenges, architecture trade-offs, and empirical benchmarks in Artificial Intelligence, Robotics, Deep Computer Vision, and Reinforcement Learning.

👨‍💻 Who I Am​

I am an AI and Robotics Software Engineer passionate about transforming theoretical algorithms into robust, real-time edge and server systems. My work spans:

  • Autonomous Systems & Reinforcement Learning: Formulating Markov Decision Processes (MDP), continuous-to-discrete state-action mappings, dense reward shaping, and Sim-to-Real hardware transfer.
  • Deep Computer Vision: Real-time multi-stage inference pipelines, object detection (YOLOv3/v5/v8), multi-object tracking (CSRT, DeepSORT), deep super-resolution (EDSR, FSRCNN), and optical character recognition (OCR) with temporal filtering.
  • Natural Language Processing & Safety AI: Parameter-Efficient Fine-Tuning (PEFT/LoRA) on Llama architectures, ordinal regression under extreme class imbalance, weak-label calibration, and deterministic guardrail wrappers.
  • Full-Stack ML Engineering: High-throughput FastAPI microservices, containerization, Docusaurus multi-lingual documentation, and cross-platform mobile interfaces (React Native).

📚 Core Series on This Blog​

1. 🧠 Maeumjigi Capstone Dev Log (23 Episodes)​

A comprehensive chronicle of developing an AI-driven crisis detection and suicide prevention system:

  • Synthetic clinical dialogue generation using large-scale LLM pipelines.
  • Overcoming severe weak-labeling overconfidence (p=0.9998) through ordinal regression and cumulative thresholding (P >= Severe).
  • Designing deterministic Safety Wrapper guardrails to guarantee sub-millisecond crisis intervention.
  • Database schema migration, connection pooling troubleshooting, and live production presentation.

2. 🤖 Webots Q-Learning Autonomous Mobile Robot (5 Episodes)​

From zero-knowledge exploration to deterministic obstacle avoidance:

  • Discretizing 6-sector continuous distance sensors into 648 compact states.
  • Counteracting sparse reward starvation with geometric velocity and distance-delta reward shaping.
  • Escaping 12-obstacle local minima traps through calibrated exploration decay ($\epsilon$-decay).
  • Deploying the learned Q-table onto physical dual-motor differential drive robots via Arduino Uno.

3. 🚗 Blackbox ANPR & Lane Intrusion Vision AI (5 Episodes)​

Real-time embedded vision system tackling high-speed highway dashcam footage:

  • 6-Stage end-to-end pipeline with strict 30 FPS latency constraints.
  • Dynamic lane polynomial fitting and geometric IoU intrusion criteria.
  • Hierarchical vehicle-to-plate localization using YOLOv3, YOLOv5, and CSRT optical tracking.
  • Restoring blurry, motion-degraded plate patches using Deep Super-Resolution (EDSR) and multi-frame stacking.
  • Korean license plate OCR engine with regex syntax normalization and 11-step temporal majority voting.

🛠️ Lab & Tech Stack​

  • Deep Learning Frameworks: PyTorch, Hugging Face Transformers, Torchvision, ONNX Runtime.
  • Computer Vision: OpenCV, Albumentations, EasyOCR, Tesseract, EDSR / FSRCNN.
  • Robotics & Simulation: Webots Robot Simulator, Arduino IDE, ROS / ROS2 fundamentals.
  • Languages & Backend: Python 3.12, TypeScript, C++, FastAPI, PostgreSQL, Redis, Node.js.
  • Hardware Rig: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM), Dual-Core ESP32 / Arduino Microcontrollers, Edge Cameras.

🌐 Open Source & Community​

All technical series published on this blog are backed by real, executable repositories and codebases. I believe that documenting failures, dead ends, and troubleshooting steps is far more valuable to fellow engineers than presenting polished final products without context.