Brennon WilliamsStudio of OneSoftware · Reasoning · Hardware

Notes from the lab · Nº 03

I wonder what the machines think.

Brennon Williams — software architect working at the seam between machine reasoning, local-first systems, and physical computing. I build production agents, dev tooling, and the occasional circuit board, mostly from a small studio with one dog and a soldering iron.

Selected work Recent writing

13 projectsactive & archive

15+ yearsshipping software

1 patentmaterial sensing probe

↳ SF · NYC · Dallas 38.89° N · 77.03° W

FIG. 01 / BW-26 Plate Nº 03 SHA · e5f884e Composed in graphite Brennon Williams, portrait

Point of ViewPrinciples I build by

The project-based life.

I have always learned by building. At fourteen I was programming 3D scanners at a medical device startup to create prosthetic fittings. Soon I was at the IDEO toy lab, designing circuits for prototypes headed to Sesame Street and Mattel. A material sensing probe I designed was patented before I was old enough to vote. None of it felt like a career — I just kept finding the edge of what I understood and building past it.

I studied Computer Science and English in college — I couldn't choose, and the two halves kept showing up in each other's work. My CS concentration was machine learning, and freshman year I built my first Hopfield network and fed it corrupted rings. Watching it reconstruct a perfect ring from a noisy, fragmentary input was my first contact with the magic of AI — the moment a system stopped feeling like code and started feeling like something mysterious and sublime. For years before ChatGPT I was building a utility-based AI framework in C# — software agents weighing competing dimensions of context to arrive at a decision.

When ChatGPT released I made a decision I'd call intentional if it hadn't also felt inevitable — go all in on large language models as the center of my work and my life. The same obsession with artificial minds, now with far greater urgency and a clearer mission. I'm an AI architect now. I help enterprise teams ship production systems — agents that make business decisions, pipelines that run at scale, orchestration that coordinates context across domains. I also train engineers on using AI in software development. After hours I keep building: a Claude Code plugin that keeps becoming a project manager, an offline voice assistant, a system for agentic long-lived modeling in Blender. It's always been the same impulse: discover something I don't understand. Build something until I do.

Passion Projects

AI & Dev Tooling

little-loops

A Claude Code plugin that turns Claude into a full development workflow engine — issue tracking, sprint planning, and FSM-based autonomous loops that drive multi-step development tasks. Ships with 21 commands and 8 specialized agents, published on PyPI.

View project Website ↗

№01

blender-agents

A production AI automation platform that runs Claude agents inside Blender for complex 3D tasks driven from natural language. A structured OODA loop autonomously executes 245 operations across 25 capability domains — modeling, rigging, materials, geometry nodes, and more.

GitHub

№02

lmc-voice

A production-grade, self-hosted real-time voice AI assistant. WebRTC audio in, streaming LLM, and tool execution across four provider categories — runs fully offline or against cloud providers, with enterprise monitoring across the stack.

GitHub

№03

deep-codebase

A CLI tool that uses Tree-sitter parsing and Claude AI to extract working implementations and map architectures from any open-source codebase. Produces byte-accurate snippets, dependency chains, and integration notes — built for developers who need code, not documentation.

Site

№04

mission-control

The central coordination repository for home network, homelab, cloud services, automation, and dev environment configuration. Single source of truth that makes the entire stack reproducible and auditable.

GitHub

№05

MC-Vault

A custom Obsidian knowledge base built on a hybrid LYT/PARA structure. 39+ automation scripts and 11 background macOS services continuously sync the vault with GitHub activity, Raindrop bookmarks, and Claude Code sessions.

№06

dotfiles

A cross-platform development environment framework managed with GNU Stow — every tool from terminal emulator to shell prompt is tuned together with a unified Gruvbox theme, conflict-resolving install, and a sub-50ms shell startup target.

GitHub

№07

swan

A lean Python multi-agent framework (~1,200 lines) that scales from a single CLI process to distributed Celery workers without changing agent code. BAML-typed LLM contracts replace stringly-typed outputs, and Microsandbox isolates execution.

GitHub

№08

graphrag-claude-code

A fork of Microsoft's GraphRAG that swaps the OpenAI API for the Claude CLI as the model backend, with local embeddings via Ollama. Integrates as a pure adapter through GraphRAG's official ModelFactory API — no core framework code modified.

GitHub

№09

Blubry

A smartphone-to-hardware bridge that turns a standard 3.5mm TRRS headphone jack into a bidirectional communication channel between phones and custom electronics. The phone's own audio codec, FFT, and mic input replace Bluetooth or USB drivers for rapid hardware prototyping.

№10

UntieAI

A C# Utility AI framework for Unity that replaces behavior trees and FSMs with a composable scoring pipeline. Pairs with an Influence Mapping layer for spatial reasoning so agents weigh regions of the world, not just direct object relationships.

GitHub

№11

UntieAnimation

A Unity animation framework that decouples AI decisions from animation logic. The agent emits named actions; physics-aware blending, predictive transitions, and editor-driven configuration handle everything that follows.

GitHub

№12

claude-loop

An early automated Claude Code agent inspired by self-directing AI research, exploring whether Claude could autonomously loop on a task without step-by-step guidance. The direct precursor to little-loops.

GitHub

All projects →

Long-form, mostly about how models think.

2025
APR

Self-Improving Reasoning LLMs through Priming: Why it Works for Some Models but Not Others

Applying Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs to LLM Development

2024
DEC

Optimizing Local LLM SVG Code Generation with TextGrad

An exploration of using TextGrad to improve SVG generation with local LLMs

All writing →

What I'm reading and recommend.

Book

Infinite Jest

David Foster Wallace

Book

Memories, Dreams, Reflections

Carl Gustav Jung

Paper

The Bitter Lesson

Rich Sutton · 2019

Paper

Building Effective AI Agents

Anthropic · 2025

All reading →

Drop me a line

brennon@brennonw.com Say hi