Skip to content

v0.1.0 - Initial Release

Latest

Choose a tag to compare

@twiecki twiecki released this 12 May 21:32

Alchemize v0.1.0

First public release of Alchemize — an AI agent that transpiles between probabilistic programming languages and compiles models to optimized native code.

pip install alchemize-ai

What it does

Alchemize uses Claude Code as a compiler. You give it a model in one framework, it autonomously writes, validates, and iterates on code in another framework — driven by gradient-distance feedback against a reference implementation. No human writes the output code.

Features

PyMC → Rust compiler

Compile any PyMC model to an optimized Rust binary, validated against PyMC's exact numerical output. 3-7x faster than the already-fast Numba backend in benchmarks across Normal, Linear Regression, Hierarchical, GP (ExpQuad), and ZeroSumNormal models.

import pymc as pm
from alchemize import compile_model

with pm.Model() as model:
    mu = pm.Normal("mu", 0, 10)
    sigma = pm.HalfNormal("sigma", 5)
    y = pm.Normal("y", mu=mu, sigma=sigma, observed=data)

result = compile_model(model)

Stan → PyMC transpiler

Translate Stan models to idiomatic PyMC, validated against BridgeStan gradients. Produced 108 out of 120 posteriordb models in a single run.

Stan → Rust compiler

Same Rust target, Stan source.

JAX ↔ PyTorch translator

Bidirectional translation between JAX and PyTorch models.

PyTorch → Rust

PyTorch models to standalone Rust binaries (includes minGPT example).

Architecture

  • Agent-based: Claude Agent SDK with tools for write / build / validate / iterate
  • Skills: Markdown files teach the agent specialized knowledge (GP CPU/CUDA/Accelerate, ZeroSumNormal, Stan→PyMC, etc.)
  • Validation loop: Every output is numerically verified against the reference implementation
  • Composable: Add a new model type or hardware backend by writing a markdown skill

CLI

alchemize convert model.py --target rust

Requirements

  • Python ≥ 3.11
  • Anthropic API key
  • Rust toolchain (for Rust targets)
  • BridgeStan + Stan toolchain (for Stan sources)

Links