Uses 持ち物 工具箱

What's in my bag

The languages, libraries and daily drivers I actually reach for, read off what my code is built with. Think of it as my inventory screen.

Languages

  • Python
  • C++
  • CUDA
  • TypeScript
  • JavaScript
  • Verilog
  • PowerShell

Machine learning

  • PyTorch
  • Hugging Face
  • Weights & Biases
  • NumPy
  • OpenCV
  • YOLO
  • CLIP
  • Kaggle

Systems & graphics

  • NVIDIA GPUs
  • CMake
  • pybind11
  • OpenSSL
  • OpenGL
  • SDL3
  • p5.js

Web & apps

  • Next.js
  • React
  • Node.js
  • Express
  • Hono
  • Bun
  • FastAPI
  • Flask
  • MongoDB
  • MySQL
  • SQLite
  • Firebase
  • Socket.IO
  • Qt / PyQt6
  • Vercel

Daily drivers

  • Arch Linux
  • Windows
  • Neovim
  • fish
  • kitty
  • Git
  • GitHub

Hardware

  • AMD Ryzen 5 5500
  • NVIDIA RTX 3050
  • 16 GB RAM
  • Kingston NVMe + SATA SSDs
  • MSI motherboard
  • Rented NVIDIA H100s for pretraining

Know-how 知識 知识

What's in my head

The ideas and techniques I work with every day, from training runs to logic gates.

AI systems

  • FSDP & ZeRO sharding
  • tensor, pipeline & expert parallelism
  • activation checkpointing
  • mixed precision (BF16, FP8)
  • FlashAttention
  • kernel fusion
  • Triton
  • CUDA graphs
  • NCCL collectives
  • KV cache & paged attention
  • continuous batching
  • speculative decoding
  • quantization (GPTQ, AWQ, INT4)
  • LoRA & QLoRA
  • roofline analysis
  • Nsight profiling

Hardware–software codesign

  • systolic arrays
  • dataflow architectures
  • memory hierarchy (HBM → SRAM → registers)
  • tiling & data reuse
  • arithmetic intensity
  • RTL design in Verilog
  • FPGA synthesis & timing closure
  • high-level synthesis
  • fixed-point arithmetic
  • structured sparsity
  • near-memory compute
  • RISC-V

Machine learning

  • backprop from scratch
  • AdamW & LR schedules
  • scaling laws
  • MoE routing & load balancing
  • Multi-head Latent Attention
  • RoPE
  • RMSNorm
  • SFT
  • RLHF
  • DPO & GRPO
  • knowledge distillation
  • evaluation harnesses

Language

  • BPE tokenizers
  • instruction tuning
  • retrieval-augmented generation
  • long-context attention
  • chain-of-thought reasoning
  • embeddings & vector search
  • MMLU, HellaSwag & ARC

Diffusion & generation

  • DDPM & DDIM samplers
  • score matching
  • flow matching
  • latent diffusion
  • classifier-free guidance
  • diffusion transformers (DiT)
  • VAEs
  • ControlNet
  • LoRA fine-tunes
  • CLIP guidance
  • video diffusion
  • text-to-speech

World models & robotics

  • latent dynamics (Dreamer)
  • JEPA
  • video prediction
  • model-based RL
  • PPO & SAC
  • imitation learning
  • diffusion policies
  • vision-language-action models
  • sim-to-real & domain randomization
  • SLAM
  • inverse kinematics
  • model predictive control
  • ROS 2