I build web-native machine learning systems and research tooling for spatial reasoning, embodied agents, and geometric representations — with a bias toward fast, reproducible experiments and clean system design.
- 🌐 Portfolio: https://moody.mx
- 🔬 Focus: web-scale RL/GA environments, high-dimensional geometry, GPU pipelines, representation learning
- 🧰 Tooling-first: I like projects that ship as usable frameworks, not just one-off demos
Core themes
- Web ML & GPU systems: WebGL / Three.js / GLSL pipelines that make training + visualization practical in-browser
- Embodied / interactive learning: Gym-style environments, instrumentation, evaluation harnesses, reproducible benchmarks
- Geometry & representations: rotations/projections in high dimensions, structured latent spaces, geometric primitives
- Gradient-free optimization: genetic algorithms, hybrid search, and alternatives/adjuncts to backprop in certain regimes
A framework for high-dimensional Snake where agents operate in continuous space with geometric action parameterizations (directional + angular components). Designed to run and visualize efficiently in the browser.
Highlights
- Web-first training/visualization pipeline (Three.js + WebGL)
- Experiments with GA / gradientless optimization and augmentation strategies
- Work on rotation-plane actions and signed-angle computations in high-D spaces
Repo: https://github.com/blayyyyyk/snake-ml
A Gymnasium-registered environment for Mario Kart DS with emulator-backed state extraction and agent evaluation.
Highlights
- Emulator instrumentation + structured observation/action interfaces
- Emphasis on making a real environment others can plug into (training scripts, eval, docs)
Repo: https://github.com/blayyyyyk/gym-mkds
Repo: https://github.com/blayyyyyk/marIOkart
Repo: https://github.com/blayyyyyk/py-desmume-mkds
Systems work to reduce GPU/CPU sync overhead and improve throughput for interactive ML visualizations on the web.
Highlights
- Practical performance engineering in WebGL pipelines
- Resource lifecycle, texture transport, and minimizing stalls
Repo: https://github.com/blayyyyyk/snake-ml
Explorations around fractional differentiation / “fractional gradients”, Fourier-domain operators, and optimization variants for neural nets (prototype-heavy, theory-curious).
A representation system for modeling LEGO constructions as a piece-graph with explicit connection semantics and robustness over connection types.
A tokenizer-inspired algorithm that mimics Byte Pair Encoding, but for encoding contours of image segments rather than text — exploring more meaningful object-level representations.
- C/C++ ⇄ Python bindings: clang-based header-to-ctypes workflows, memory mapping + structured state extraction
- TensorFlow.js training loops: scope/lifetime management, GPU memory behavior, stable in-browser trainers
- Three.js internals: textures, uniforms, WebGL resource binding, performance pitfalls
- Dataset utilities: memmap pipelines, synthetic DB generation, and reproducible data builds
If you’re browsing my repos, check:
- pinned projects (curated “mainline” work)
Languages
- Python, TypeScript/JavaScript, C/C++, GLSL, (plus occasional Racket for theory coursework)
ML / RL
- PyTorch, TensorFlow.js, Gymnasium, custom GA harnesses
Graphics / Web
- WebGL, Three.js, React/Next.js, shader tooling
Infra / dev
- WSL/Linux tooling, performance profiling, reproducible scripts
- Snake-ML: manuscript + revisions in progress (web-native ML for N-D reasoning tasks)
- MarI/O Kart: paper-track framing in progress (environment + agent evaluation)
- Website: https://www.moody.mx/
- GitHub:
@blayyyyyk - Email:
blake@mpsych.org
I’m most interested in collaborations around:
- web-deployable ML systems
- embodied agent benchmarks + environments
- geometric representations / structured priors
- performance engineering for interactive ML