interactive.paiml.com is the home of Pragmatic AI Labs' WASM-powered books and demos — and a window into the four open-source projects we use to build production AI without dependencies, cloud APIs, or vendor lock-in.
aprender trains and serves models. forjar provisions the hardware. xpile brings legacy code along for the ride. pmat proves it all works.
Next-generation ML in pure Rust. One install, one binary, full stack.
57 commands across inference, training, model ops, quantization, profiling, registry, and GPU — backed by 405 equation-based contracts.
apr run · chat · serve · run GGUF, SafeTensors, .apr models locallyapr finetune · train · distill · LoRA / QLoRA fine-tuningRust-native IaC. Bare-metal first. BLAKE3 content-addressed state.
One binary, 17 deps, zero cloud APIs. Drift detection via local hash compare. Apply in milliseconds, not minutes.
forjar plan · apply · drift · YAML → DAG → ssh executionmodel, gpu, pepita sandboxBring any codebase to Rust. Python, C, and a new language built for it.
Three transpilers and a programming language, sharing one philosophy: ownership-aware codegen, formal verification, and energy-efficient output.
depyler · Python → Rust · semantic verification, MCP integrationdecy · C → Rust · ownership inference, 0 unsafe blocks by defaultruchy · self-hosting language · transpiles to Rust, beats C at runtimeMCP agent toolkit. Make code-with-agents deterministic.
Six orthogonal quality metrics (TDG), mutation testing, WASM formal verification, and zero-config context generation for any codebase.
pmat context · LLM-optimized repo context for Claude Code, Clinepmat analyze tdg · A+ → F grade across complexity, SATD, coverage…pmat analyze wasm --verify · formal proofs of memory safetyWeb OS, .apr model playground & kernel microVM. Real aprender ML models with 200+ Popperian falsification tests. Pure Rust/WASM microkernel.
96 years of S&P 500 data, simulated live. SIMD-accelerated with aprender + trueno.
WASM-powered interactive book. TDD-first, implementation-driven. Beats C on the hello.ruchy benchmark.
Neural net opponent · jugar engine · 100% Rust/WASM.
N-body physics in real time · simular engine.
Genetic algorithms · interactive city placement.
Train/inference from the command line with aprender.
.apr format configuration · presets + custom settings.
Browser-native ML model inspector · ZSTD .apr loader.
PMAT quality metrics, formal verification. Interactive presentation.
Click, argparse, professional CLI development patterns.
Python fundamentals with a minimalist approach.
pytest, unittest, TDD · build with confidence.
NumPy, Pandas, real datasets · interactive throughout.