# lulzx projects

Small, focused software projects by lulzx. Most are open source and built to do one job well. Ordered by depth of the work rather than by date. This list is the site root; the rest of the site starts at [lulzx.com/landing](https://lulzx.com/landing).

- [cuda-metal](https://github.com/Lulzx/cuda-metal): CUDA compiler and runtime for Apple Silicon: it compiles and runs CUDA kernels on the M-series GPU.
- [gromacs-metal](https://github.com/Lulzx/gromacs-metal): full-GPU PME for GROMACS on Apple Silicon through AdaptiveCpp's SYCL-to-Metal backend, with a VkFFT Metal bridge.
- [zpdf](https://github.com/Lulzx/zpdf): zero-copy PDF text extraction in Zig. Memory-mapped parsing with SIMD acceleration.
- [warpx-metal](https://github.com/Lulzx/warpx-metal): WarpX electromagnetic particle-in-cell simulations on the Apple Silicon GPU through an AdaptiveCpp SYCL Metal backend.
- [hvm](https://github.com/Lulzx/hvm): a Zig implementation of HVM, the Higher-Order Virtual Machine based on the interaction calculus.
- [arturo-compiler](https://github.com/Lulzx/arturo-compiler): a self-hosting compiler for Arturo, written in Arturo.
- [tacitj](https://github.com/Lulzx/tacitj): a self-hosting compiler for a tacit-leaning subset of J, written in J.
- [WrenJIT](https://github.com/Lulzx/WrenJIT): a tracing JIT compiler for the Wren scripting language.
- [VF64-metal](https://github.com/Lulzx/VF64-metal): a virtual FP64 research harness for Apple Metal GPUs, which have no hardware double precision.
- [tiny-kernel](https://github.com/Lulzx/tiny-kernel): fused Metal kernels for LLM inference, written in Zig. No Python, no PyTorch.
- [zeno](https://github.com/Lulzx/zeno): GPU-accelerated batched physics simulation for robotics and reinforcement learning on Apple Silicon.
- [lulang](https://github.com/Lulzx/lulang): a numerics-first programming language.
- [sectorc](https://github.com/Lulzx/sectorc): a C compiler built from first principles, on the premise that you should trust no compiler you have not verified.
- [tacit](https://github.com/Lulzx/tacit): an array-transformation OS written in Uiua. It schedules graphs, not threads.
- [beam.zig](https://github.com/Lulzx/beam.zig): a minimalist BEAM virtual machine in Zig.
- [zig-tvm](https://github.com/Lulzx/zig-tvm): a compact, table-driven implementation of the TON Virtual Machine in Zig.
- [wasm-mquickjs](https://github.com/Lulzx/wasm-mquickjs): MicroQuickJS compiled to WebAssembly.
- [zs3](https://github.com/Lulzx/zs3): S3-compatible object storage written in Zig with zero dependencies.
- [tinypdf](https://github.com/Lulzx/tinypdf): a PDF creation library in less than 400 lines of TypeScript. No dependencies, and the files it writes are real PDFs.
- [blaze-bpe](https://github.com/Lulzx/blaze-bpe): a lightweight embeddable tiktoken-compatible BPE tokenizer core in Rust.
- [supsearch](https://github.com/Lulzx/supsearch): a program synthesis engine that solves every problem in LamBench with no model and no tokens, certified by the benchmark's own checker.
- [ontogenesis](https://github.com/Lulzx/ontogenesis): a lambda-calculus synthesizer that invents its own vocabulary as it searches.
- [progen](https://github.com/Lulzx/progen): program synthesis through HVM4 superpositions.
- [subquadratic-sparse-attention](https://github.com/Lulzx/subquadratic-sparse-attention): an MLX prototype for content-routed subquadratic sparse attention and long-context associative recall.
- [lulzfish](https://github.com/Lulzx/lulzfish): a high-performance chess engine.
- [barunlm-35m-browser](https://github.com/Lulzx/barunlm-35m-browser): the BarunLM-35M language model converted to ONNX and run entirely client-side with onnxruntime-web.
- [nn.zig](https://github.com/Lulzx/nn.zig): a neural network in roughly 600 lines of Zig, training with GEMM batching through Accelerate.
- [nabla](https://github.com/Lulzx/nabla): reverse-mode automatic differentiation on NumPy: exact gradients through any composition.
- [solon](https://github.com/Lulzx/solon): a language learner built on minimum description length. No transformer and no backpropagation.
- [helicon](https://github.com/Lulzx/helicon): GPU-accelerated magnetic nozzle simulation and plasma detachment analysis.
- [driftmap](https://github.com/Lulzx/driftmap): vortex flow map methods for simulating plasma edge turbulence.
- [skein](https://github.com/Lulzx/skein): a GPS-denied drone-swarm autonomy simulator with EKF localisation, cooperative SLAM and gossip consensus.
- [microduck-backflip](https://github.com/Lulzx/microduck-backflip): a reproducible reinforcement-learning task, evaluation pipeline and safety gates for teaching a Pollen Robotics MicroDuck to backflip.
- [zsearch](https://github.com/Lulzx/zsearch): a branchable, content-addressed full-text search index in Zig.
- [walsync](https://github.com/Lulzx/walsync): streams SQLite's write-ahead log to S3 and restores it on demand, so a database can follow you between machines.
- [GeoPERSIST](https://github.com/Lulzx/GeoPERSIST): persistent geospatial world models.
- [soma](https://github.com/Lulzx/soma): a small-scale SIMD operating system for many-agent systems.
- [typed-inference](https://github.com/Lulzx/typed-inference): a schema-compiled grammar runtime that guarantees structured output from a language model.
- [recall](https://github.com/Lulzx/recall): RECALL-500, a benchmark measuring where retrieval-augmented generation fails against differential memory.
- [memo](https://github.com/Lulzx/memo): permanent memory for agents in 150 lines of Python: one log file and one projection.
- [caas](https://github.com/Lulzx/caas): others persist computation. This persists responsibility: promises kept even when the promiser is gone.
- [caplock](https://github.com/Lulzx/caplock): stops dependencies from reaching anything they were not explicitly granted.
- [encept](https://github.com/Lulzx/encept): perceptual image hashing that uses Apple's hardware H.264 encoder as a learned similarity function.
- [paragon](https://github.com/Lulzx/paragon): provably correct model training on Bend's massively parallel runtime.
- [reality-fit](https://github.com/Lulzx/reality-fit): a webcam scan and a printed marker produce a 3D-printed part that fits on the first try.
- [ego-death](https://github.com/Lulzx/ego-death): a mathematical framework modelling psychedelics as 2-cell transformations in predictive coding hierarchies.
- [pybin](https://github.com/Lulzx/pybin): ships a Python application as a single executable.
- [mirror](https://github.com/Lulzx/mirror): bidirectional functions for JavaScript. One function, both directions, no duplication.
- [tinydocx](https://github.com/Lulzx/tinydocx): a minimal DOCX and ODT creation library with no dependencies.
- [tinysoup](https://github.com/Lulzx/tinysoup): a zero-dependency HTML parser for TypeScript with a BeautifulSoup-like API.
- [tinybarcode](https://github.com/Lulzx/tinybarcode): SVG barcode generation in under 1KB gzipped.
- [reverse-geocode](https://github.com/Lulzx/reverse-geocode): offline reverse geocoding in the browser, comparing a Morton grid against two H3 approaches.
- [press](https://github.com/Lulzx/press): drop in a video and get a small one back. A native macOS compressor built with Zig and ffmpeg.
- [ubrowser](https://github.com/Lulzx/ubrowser): fast and cheap browser automation for Claude Code.
- [carry-terminal](https://github.com/Lulzx/carry-terminal): an agent-native terminal, written in Rust.
- [mimo-build](https://github.com/Lulzx/mimo-build): a terminal AI coding agent written in Rust.
- [lulz-router](https://github.com/Lulzx/lulz-router): runs Claude Code, Codex or OpenCode on one subscription, bridging the Responses and Chat Completions APIs.
- [open-tag](https://github.com/Lulzx/open-tag): a model- and platform-agnostic alternative to Claude Tag: a persistent AI teammate for Telegram and Discord.
- [social-stockfish](https://github.com/Lulzx/social-stockfish): Monte Carlo tree search over conversations: it looks several moves ahead to suggest what to say.
- [rat-runner](https://github.com/Lulzx/rat-runner): a third-person rat sandbox in three.js, with shell fur through alpha-to-coverage.
- [life-sim](https://github.com/Lulzx/life-sim): an interactive multiscale WebGL simulation of a living E. coli cell.
- [living-in-a-van](https://github.com/Lulzx/living-in-a-van): a five-minute animated short rendered from a single HTML file, with hand-drawn SVG and local text-to-speech voices.
- [flute](https://github.com/Lulzx/flute): sixteen interfaces that only make sense above a million tokens per second.
- [rockylm](https://github.com/Lulzx/rockylm): a 9M-parameter language model that talks like Rocky from Project Hail Mary, with speech in his cloned voice.
- [telekit](https://github.com/Lulzx/telekit): native-feeling UI components for Telegram Mini Apps. No dependencies and no build step.
- [esp32-cam-web-streamer](https://github.com/Lulzx/esp32-cam-web-streamer): live MJPEG webcam firmware for the AI-Thinker ESP32-CAM, viewable in any browser over WiFi.
- [latent-sokoban](https://github.com/Lulzx/latent-sokoban): a pixel-based latent world model competition.

Everything else is on [github.com/Lulzx](https://github.com/Lulzx).

There is no site-level API for project data. Follow each repository for source, releases, issues, and project-specific documentation.