Building efficient learning systems — from language models to low-latency robot control.
- 🎓 Konkuk University — Computer Science & Engineering
- 🔬 Interested in Reinforcement Learning, NLP, Model Compression, and Physical AI
- 🧩 I enjoy turning research ideas into reproducible experiments and working systems
- 🌱 Currently studying policy distillation and compression for real-time robot control
A research pipeline for distilling a high-performance PPO teacher into a compact student policy and measuring real CPU latency gains from quantization and structured pruning.
| Stage | Status |
|---|---|
| Minimal PPO baseline | ✅ Completed — Reacher-v5, 2.5M environment steps |
| Reproducible evaluation | ✅ Completed — 30 seeded episodes |
| Improved PPO teacher | 🚧 In progress |
| Teacher–student distillation | 🧱 Pipeline scaffolded |
| Quantization & structured pruning | 🧱 Evaluation interfaces scaffolded |
Research focus: PPO · Knowledge Distillation · PTQ/QAT · Structured Pruning · CPU Latency Benchmarking
Implemented an encoder–decoder Transformer directly in PyTorch and compared full fine-tuning with LoRA ranks 4, 8, and 16 on approximately 1.6M Korean–English sentence pairs.
- Built the Transformer architecture and training pipeline from scratch
- Evaluated with BLEU, chrF, BERTScore, perplexity, throughput, and GPU memory
- LoRA rank 8 trained only 0.4842% of total parameters
- Reduced training time by about 9.14% and peak GPU memory by about 20% compared with full fine-tuning
PyTorch Transformer LoRA SentencePiece NLP Parameter-Efficient Fine-Tuning
✈️ TripTailor
AI-powered personalized travel curation platform developed as a five-person team project.
- Recommends destinations using user preferences, tags, and vector search
- Supports itinerary creation, sharing, reviews, and recommendation feedback loops
- Contributed to AI recommendation features and frontend integration
- Built and deployed as a Django service with Docker, Gunicorn, Nginx, and GitHub Actions
Django Python JavaScript Vector Search AI Recommendation Docker
AI / Research
Development
Open to conversations about AI research, efficient models, and collaborative projects.