evolve.h
C++26 evolutionary algorithm framework. Bandit-driven operator selection, multi-island topology with migration, and adaptive mutation rates. Currently chasing Schur(6) via large-scale combinatorial search.
I have always been obsessed with bounding problems. Finding the edges and limits of things, whether performance, functions, or axiomatic systems. The idea that these things have shapes is endlessly interesting to me. This has taken me across an enormous range of interests, but the ones you may know me for are RL, trading research, and software engineering. At work I focus on semi-Markov systems and microstructure; outside of it, deep RL and writing maximally performant systems. Right now I am particularly interested in the intersection of compiler design and reinforcement learning — generating optimal bytecode sequences via policy search.
I don't do this for compensation; doing these things well is the compensation. I love it, and I always have. I have written software and researched at many familiar firms: Tower Research Capital, TRV Trading, and contributed to public research in QR many times, particularly regarding the martingale midprice. I have also historically quoted BTC-USDT on Binance using my own software and research stack.
Papers & Code
Companion note connecting the microprice to semi-Markov reinforcement learning. Derives discounted move-time Bellman equations, event-time differential values, and TD learning rules with empirical benchmarks.
A centered Poisson framework that recasts the microprice as a martingale corrector to the midprice. Derives fair-value estimators with empirical benchmarks on IEX and WSELOB data.
reinforce
Header-only C++26 reinforcement learning library. Type-safe tabular RL with compile-time hyperparameter validation, policy iteration, and TD(λ) methods.
docgen
Zig documentation generator that parses C++ headers and emits static HTML reference sites. Handles templates, concepts, and nested namespaces.
log.h
Single-header C++23 structured logging library. Zero heap allocations on the hot path, ~110ns per formatted call, SIMD-accelerated field serialisation.
Writing
mathsceling — notes on microstructure, operator methods, and whatever else converges. coming soon.