Projects

CJC-Lang — Computational Jacobian Core Beta

A deterministic programming language for scientific computing, built from scratch in Rust with nearly zero external dependencies.

CJC-Lang covers the full stack: lexer, parser, type checker, two intermediate representations (HIR and MIR), an optimizer, two execution engines, and a runtime with ~300 built-in functions spanning math, statistics, linear algebra, signal processing, machine learning, and data manipulation.

Key technical facts:

  • 6.4 MB single binary — the entire language in one file
  • 20 Rust crates, nearly zero external dependencies
  • 5,300+ passing tests across 55 test binaries
  • Deterministic execution: same seed = bit-identical results

Compiler and Runtime

Dual-Engine Compiler Pipeline

CJC-Lang has two execution paths that must always agree:

  1. AST Interpreter (cjc-eval) — direct tree-walk execution for rapid prototyping and debugging
  2. MIR Executor (cjc-mir-exec) — register-machine execution through HIR → MIR lowering with optimization

Parity tests verify that both engines produce identical results for every program. The MIR path supports constant folding and dead code elimination (--mir-opt), and a NoGC static verifier proves that annotated code paths never trigger garbage collection.

Type System and Diagnostics

Unification-based type checker with explicit annotations. Comprehensive diagnostic engine with structured error codes (E0xxx–E8xxx) and rich terminal output. Language features include closures with capture analysis, match expressions with structural destructuring, for-loops, structs, enums, and tuples.

Scientific Computing

Tensor and Linear Algebra Runtime

N-dimensional tensors with zero-copy COW (copy-on-write) buffers. Operations include matrix multiplication (sequential, tiled, and parallel), SVD, QR, Cholesky, eigenvalue decomposition, determinant, inverse, and norm computation.

All floating-point reductions use Kahan or binned summation to maintain numerical stability. SIMD kernels avoid FMA instructions to preserve cross-platform bit-identical results.

Automatic Differentiation

Forward-mode AD via dual numbers and reverse-mode AD via tape-based gradient computation. Supports standard operations (arithmetic, trigonometric, exponential, logarithmic, power) with automatic derivative propagation.

Used in practice for physics-informed neural networks (PINNs) and physics-informed machine learning (PIML) — demonstrated on heat equation and harmonic oscillator problems with verified convergence.

Statistics and Distributions

35+ statistical functions (mean, std, variance, correlation, covariance, percentile) and 10 probability distributions (normal, beta, gamma, exponential, F, Weibull, Poisson, binomial, chi-squared, Student’s t). 12 hypothesis tests (t-test, chi-squared, ANOVA, Wilcoxon, Mann-Whitney, Kruskal-Wallis). All seeded and deterministic.

Data Science

Tidyverse-Inspired Data DSL

73+ DataFrame operations modeled on R’s tidyverse: filter, select, mutate, group_by, summarize, arrange, pivot_longer, pivot_wider, joins (inner, left, right, anti, semi), and window functions (lag, lead, cumsum, ntile).

Supports typed columns (int, float, string, boolean, categorical, datetime), CSV ingestion, and lazy evaluation with COW semantics.

Benchmarks and Demonstrations

Chess RL Benchmark

A full chess environment and REINFORCE reinforcement learning agent written entirely in CJC-Lang. 69 tests covering board invariants, legal move generation, policy gradient updates, and training determinism. Validates that CJC can express a complete ML workflow end-to-end.

Physics-Informed ML (SciML)

PIML demonstration: 1D heat equation solved via polynomial regression with physics residual penalty. PINN demonstration: harmonic oscillator solved with a neural network surrogate. Both verified for solution accuracy, PDE residual convergence, gradient correctness, boundary condition satisfaction, and bit-identical determinism across runs.

Tooling

CLI Toolchain

15+ subcommands: run, check, parse, lex, repl, plus inspection (inspect, schema), analysis (trace, flow, drift), performance (bench, mem), and utility (pack, doctor, forge, patch). Supports --reproducible mode, --seed control, --mir-opt optimization, and rich diagnostic formatting.

Visualization (Vizor)

Grammar-of-graphics visualization library built into the CJC-Lang ecosystem. Declarative plot specification following principles similar to ggplot2. Currently integrated at the library level within the CJC-Lang runtime.