Zefan Cai caizefan34

👨‍💻 About Me

I'm a university student & open-source researcher passionate at the intersection of AI, Reinforcement Learning, and 3D Computer Graphics.

🔬 Building research-grade AI benchmarks with honest, reproducible evaluation
🚀 Exploring 3D Gaussian Splatting (3DGS) optimization and urban mobility AI
📊 Advocate for reproducible research — strict train/test splits, checked-in reference metrics, and honest negative results
🎯 Believer in rigorous experimentation over hype-driven claims

AI-driven taxi fleet repositioning on 263 NYC zones — demand forecasting, multi-agent simulation, offline RL & reproducible OPE

	Highlights
🧠	Forecasting — LightGBM, XGBoost, Ensemble, GraphSAGE
🤖	RL Policies — DQN, Double DQN, IQL with paired statistical tests
🚦	Multi-Agent Simulator — calibrated competition, demand depletion
✅	402 passing tests, reproducibility badges, strict temporal splits
🏅	NDCG@3 0.9565 · +$54/driver DQN lift (p < 1e-10)
🌐	Live Interactive Demo

Which 3DGS renderer is fastest? Which compression is best? We measured them all — on the same GPU, same scenes, same protocol.

	Highlights
⚡	12 renderer variants — gsplat, 3DGS, Speedy-Splat, TC-GS, and HiGS
📐	5 real-world scenes × multi-metric: FPS, PSNR, SSIM, LPIPS, VRAM
🔬	HiGS ablation — 7× sparse-training speedup research with quality retention
📉	Pareto-optimal analysis — identify true win-win trade-offs
🌐	Live Results

A private, local AI workspace for RAG, knowledge bases, coding help, and LoRA fine-tuning.

	Highlights
🧩	Ollama + FastGPT + pgvector — full local AI stack
📚	RAG pipeline — document retrieval, chunking, embeddings
🎯	LoRA fine-tuning — customize models on your own data
🐳	Docker Compose — one-command deployment

"Reproducibility is not optional."

Every benchmark in my repositories includes:

_{⚡ Built with rigor. Measured with honesty. Shared with purpose.}