I build the layer underneath before I use it — a CNN library in pure NumPy with its own optimizers and metrics, a 3D renderer with its own linear algebra, panorama stitching from SIFT and RANSAC up. Then I measure what I built: controlled experiments, held-out seeds, and the distribution rather than the average.
Two single-author papers, each in its repo:
Multi-Agent Financial Market Laboratory — five markets, identical except for how many momentum traders live in them. Fragility doesn't grow smoothly; past a threshold the market changes regime. Median PnL is negative in every configuration while the mean stays positive.
Evolutionary RL for Housing Price Volatility Control — a regulator that cannot set prices, only credit, supply and friction, and only sees the effect after a lag. PPO learns the policy; an evolutionary loop searches over what PPO is given.
Also here: SafeAgree, an LLM-ready dataset built from raw HTML privacy policies with reproducible splits — and my dotfiles: Hyprland, Emacs and Neovim, configured the way I actually work.