- 🎓 Status: First-year Statistics student at UC Berkeley.
- 🧑💻 Identity: A developer building ML systems, computer-vision pipelines, and interactive character applications.
- 🚣 Sports: A passionate Canoe Polo athlete.
- 🎯 Focus: Exploring multimodal AI, efficient LLM inference, model evaluation, and interactive systems.
- 🎮 Daily Life: Coding, watching anime, gaming, and turning ideas into reality.
Here is a simple index of the projects I am currently working on.
- 🛟 H20Saver: A PyTorch/Ultralytics YOLO11x fine-tuning pipeline for drowning-target detection, including dataset analysis, training, validation, and ONNX export.
- 🔀 airi-llm-router: A FastAPI and asyncio gateway for local LLM inference, with payload-aware queues and GPU-memory backpressure.
- 🖥️ OwO-Desktop: A Windows Live2D desktop character application with WebSocket controls, a Python LLM/TTS backend, and expression, motion, and lip-sync integration.
- 🦙 llama-cpp-moe-vram-benchmarks: An automated llama.cpp benchmark/report exploring CPU/GPU MoE offload, 4-bit KV cache, and Flash Attention on consumer hardware.
- 🌸 My-Spiritual-Solace: A collaborative Live2D dynamic model & resource sharing repository bringing favorite anime characters to life on screen.
- 🌐 Layer0-kol: Team collaboration project.
- ML & Data: Jupyter notebooks, CUDA, ComfyUI, ONNX, Pandas, NumPy, pynvml
- Native & Interactive: llama.cpp, Direct3D 11, Live2D Cubism Native SDK, IXWebSocket, nlohmann/json, Asyncio