CodeReclaimers, LLC · Principal: Alan McIntyre · Est. 2014
Independent software consulting, sole operator since 2014. Twenty-eight years shipping production code in C++, Python, and Julia; MS in Applied Mathematics from the University of Washington. Maintainer of neat-python, NeatEvolution.jl, Arborist.jl, and benchstone; operator of the Caliper and Occam MCP services; author of the adversarial-loop Claude Code skill. Current focus: agentic-environment design, evaluation infrastructure, and LLM-augmented research engineering.
I. Focus
Four overlapping areas. They share a temperament more than a technology stack: problems where the math is real, the correctness matters, and somebody has to stand behind what ships.
Geometry kernels, reverse engineering, and numerical algorithms for engineering software. Lead developer for DezignWorks (Creative Dezign Concepts, Inc.), a reverse-engineering addon suite for SolidWorks, Autodesk Inventor, and Creo Parametric, with data acquisition for Romer, Faro, Leica, and other metrology devices.
Differential equations, optimization, linear algebra, and high-performance numerical work in C++, Python, and Julia. CUDA / GPU programming for performance-sensitive workloads. Library development, code review, and implementation engagements for scientific computing teams.
Applied research and open-source stewardship. Maintainer of neat-python (cited in 100+ academic papers), NeatEvolution.jl, and Arborist.jl — a generic genetic programming framework for Julia with NEAT, NSGA-II, quality-diversity, and LLM-as-operator support.
Agentic-environment design, evaluation-harness construction, and multi-model orchestration patterns for research workflows. Builder of benchstone — a statistical-gate evaluation harness for code-optimization loops — and the adversarial-loop Claude Code skill (Claude-orchestrator + Codex/Gemini critics). Implementation, audit, and tooling engagements for teams shipping LLM-augmented research and development pipelines.
II. Selected work
A partial inventory. The open-source libraries are used by other researchers and practitioners; the MCP services run in production and take real payments; the Claude Code skill encodes a multi-model orchestration pattern used in daily work.
Python implementation of the NEAT (NeuroEvolution of Augmenting Topologies) algorithm for evolving neural network topologies and weights simultaneously. Maintained since 2015. Version 2.x introduces per-node time constants, enabling continuous-time recurrent neural network (CTRNN) evolution.
github.com/CodeReclaimers/neat-python →Julia implementation of NEAT-family neuroevolution algorithms, registered in the Julia general package registry. Designed for performance on scientific computing workloads, with native integration into the Julia SciML ecosystem.
github.com/CodeReclaimers/NeatEvolution.jl →Generic, extensible genetic programming framework for Julia. First-class support for NEAT neural topology evolution, LLM-as-operator (FunSearch / AlphaEvolve pattern), symbolic regression, multi-objective optimization (NSGA-II), quality-diversity search (Novelty Search, MAP-Elites), and CMA-ES. Follows the SciML Problem/Algorithm/Solve pattern.
github.com/CodeReclaimers/Arborist.jl →Geometry and CAD file metadata MCP server. Accepts geometry files by URL or base64 payload, extracts metadata and statistics, returns structured JSON. Handles STL, OBJ, PLY, PCD, LAS/LAZ, and glTF/GLB; STEP and IGES are on the roadmap. Per-call payment via x402 (USDC on Base) or Stripe.
caliper.fit →Symbolic regression as a service via MCP. Exposes sindy_run (sparse identification of nonlinear dynamics from time series), pysr_run (evolutionary symbolic regression via SymbolicRegression.jl, returning a Pareto front of complexity-vs-accuracy expressions), and pysr_uncertainty (bootstrap confidence intervals for fit constants). Free tier and paid tier billed via x402 or MPP/Stripe.
occam.fit →Trustworthy benchmark gating for code-optimization loops. A CLI that answers “did that change actually help?” with PROMOTE / REJECT / NEEDS_MORE_DATA verdicts, append-only SQLite history, harness-owned reference artifacts the agent cannot reach, and statistical gates (direction-aware sigma or rank-based Mann-Whitney for heavy-tailed metrics). Designed for AI-agent autoloops; works equally well without one.
github.com/CodeReclaimers/benchstone →A Claude Code skill that runs an automated adversarial review loop over a research project. Claude Code orchestrates; an external critic (Codex CLI and/or Gemini CLI, with Opus 4.7 as a fallback) generates testable hypotheses about what might be wrong; Claude Code filters and investigates the substantive ones empirically. Typical output: 5–15 actionable findings per 3-iteration batch, 40–70% confirmation rate, 20–45 minutes wall-clock per batch.
github.com/CodeReclaimers/adversarial-loop →III. About
CodeReclaimers is a one-person practice founded in 2014. I take on a small number of engagements at a time, typically in computational geometry, scientific computing, or neuroevolution — work where the math is real, correctness matters, and somebody has to stand behind what ships.
Engagements span four shapes: research consulting (algorithm design, architecture review, and implementation guidance for academic and industry research projects); library development (feature development, performance optimization, and API design for scientific C++, Python, and Julia libraries); implementation engagements (end-to-end implementation of algorithms from papers or specifications); and code review and audit (structured review of technical or scientific computing codebases, with written findings).
I use AI tools where they make the work faster or better — I'm the one responsible for what comes out the other side. Every engagement is worked by me: I read the code, I shape the solution, I sign the report.
Currently available, with a typical four-week start horizon. Engagements run anywhere from 4 to 52 weeks. A short email to consulting@codereclaimers.com describing the problem is enough to start; replies typically within two business days. The full consulting profile is also available as /llms.txt and /capabilities.json for agent consumption.
IV. Field notes
Notes from the workshop — rescued codebases, numerical methods, tools I've found useful, the occasional post-mortem. Written when there's something worth writing down.
A nanoGPT optimization loop accepted a one-line patch that quietly disabled dropout. Whatever the harness doesn’t pin, the optimizer treats as up for grabs — and the catches keep landing at the friction point, not the artifact.
Running 25 experiments over three days with Claude Code and Gemini Deep Think to optimize a sketch reconstruction pipeline, and what the rejection pattern revealed about detection-selection architectures.
How I turned 20 years of email, forum posts, and chat logs into a personality model that can write in my voice — and the recursive absurdity of using it to write about itself.
Embedding a Recursive support agent on your website is as simple as pasting an iframe snippet. A step-by-step guide covering domain allowlisting, sizing, positioning, and troubleshooting.
How a YouTube video about AI resume assistants led to building Recursive, a platform that gives small businesses self-improving AI support agents.
The alpha release of SketchBridge, a desktop application for transferring 2D sketches between CAD platforms including FreeCAD, Fusion 360, SolidWorks, and Inventor.
V. Contact
A short email describing the problem is enough to start. Replies typically within two business days. Also on GitHub and ORCID; see the résumé for the long form, or ask the AI résumé a question.