utilitas et venustas — utility & delight

WRITING

Documentation of technical experiments, architecture patterns, and thoughts on the AI industry.

Engineering |

Bridging Physical Audio to AI Agents: Building a Production Sonos MCP Server in Go

Controlling physical hardware through an LLM sounds straightforward until your agent burns 2,000 tokens on raw UPnP XML, drops commands on grouped follower speakers, or blasts volume at 100%. Built in anticipation of Sonos 27's official MCP server, here is how we engineered a resilient, zero-dependency Go MCP server for Sonos whole-home audio.

Read Story
Engineering |

How to Build a Production-Ready, Secure MCP Server in Go

Exposing MCP tools in production requires solving cold-start disconnects, unauthenticated ingress, and authorization latency. This field guide walks through building a secure, high-performance Go MCP server with dual-transport multiplexing, stateless Streamable HTTP, OAuth 2.1, and granular per-tool scopes.

Read Story
Engineering |

Grading the Agent, Not the Pixels: Spec-Driven Evaluation for Generative-Media MCP Agents

Evaluating generative-media AI agents on pixel and audio quality is a costly, slow, and defect-confounding trap. Here is how we reproduced Apple's Agent Seer framework across Google Veo, Nanobanana, and Lyria MCP servers, uncovered catastrophic Schema-Blindness in LLM judges, and fixed it with runtime capability matrices.

Read Story
Engineering |

Can Any App Become an Agent Plugin? Assessing Agent Readiness with Montage

The rush to wrap legacy applications into Model Context Protocol servers has created bloated prompts, arithmetic hallucinations, and broken workflows. Montage is a framework and diagnostic toolchain designed to answer two fundamental questions: can any application become a set of agent skills and plugins, and more importantly, should it?

Read Story
Thought experiments |

The Nine Layers of Agent Memory

Talking about "the agent memory layer" in the singular is handwaving. Agent memory is a stack of nine distinct concerns across three planes, where the formal Semantic Web past and the pragmatic markdown reboot collide. Here is the full stack, where OKF fits, and which past ideas are worth bringing back.

Read Story
Thought experiments |

The Open Knowledge Format Ecosystem: A Field Guide

A survey of the tools growing up around the Open Knowledge Format: desktop apps, agent tooling, importers, and libraries, and where the gaps still are. OKF gives portable shape to the markdown "agent exhaust" agents leave behind. Includes where binder provides the missing corpus on-ramp, and how Agent Skills and Plugins integrate OKF with AI workflows.

Read Story
Tooling |

Taming Cloud Sprawl and AI Spend: Hands-On with GCP Management Agent Skills

Over years of rapid prototyping, Google Cloud portfolios accumulate zombie disks, dormant codelabs, and unexpected GenAI token spend. Here is how to package read-only portfolio auditing, BigQuery billing velocity analysis, quota governance, and safe decommissioning into an open Agent Plugin for agy and OpenCode.

Read Story
Thought experiments |

Deciding What to Automate: An Agent Skill for DeepMind's Intelligent Delegation Framework

An open Agent Skill implementing Google DeepMind's Intelligent AI Delegation framework (arXiv:2602.11865). A 5-gate diagnostic funnel, hard-gate vetoes, and 4 automation tiers translate DeepMind's governance research into an actionable evaluation tool for workflows and processes.

Read Story
Tooling |

Best Practices with binder: When to Reach for the CLI, Skills, or MCP

A hands-on guide to binder, the tool that turns a markdown corpus (often the "agent exhaust" your agents already left behind) into a conformant OKF bundle. Start with a quickstart, then decide which surface to drive it through: the CLI, an Agent Skill, or an MCP server all ship today.

Read Story
Experiments |

Intro to Moonshine Voice Go Samples

By combining Go-native runtimes with local streaming models, `moonshine-go` bypasses cloud latency bottlenecks to deliver private, offline voice interactions in under 100 milliseconds. This showcase of 11 runnable applications demonstrates how to build deterministic, highly composable voice agents and copilots that run entirely on local hardware.

Read Story
Experiments |

Hermes Agent x A2A with Google Cloud

A reference for a Google Cloud project-scoped deployment of the Nous Research Hermes Agent integrated with Google Cloud AI. This project showcases robust implementation of bidirectional Agent-to-Agent (A2A) communication, secure command sandboxing, and deployment-ready topologies.

Read Story
Experiments |

Compressing OpenCode's Google Cloud AI Traffic with Headroom

Learn how to integrate Headroom and OpenCode with Vertex AI to compress LLM traffic for Claude, Gemini, and Gemma 4 on Google Cloud. This technical walkthrough demonstrates how to slash token consumption by up to 90% while maintaining model accuracy.

Read Story
Experiments |

jiborelay: A Robot That Only Knew localhost

jiborelay is a lightweight Go tool designed to bypass Jibo's hardcoded firmware limitations by forwarding its local network traffic to external backends. By quietly routing local bytes to modern cloud APIs, it untethers the vintage robot's brain from its decade-old hardware and breathes new life into a beloved social robot.

Read Story
Engineering |

Jibo’s Independence Day: Unlocking a Social Robot

By reverse-engineering the social robot's APIs and utilizing community hardware exploits, I built a custom Go-based cloud replacement to revive a disabled Jibo after nine years on a shelf. Integrating Gemini for real-time conversation and tool-calling, this project successfully bridges legacy robotics with modern agentic AI.

Read Story
Engineering |

Frame by Frame: Lessons Building on Magenta RealTime 2

With Google DeepMind's Magenta RealTime 2, I built native SwiftUI and Rust players that generate continuous, real-time music locally on Apple Silicon. This project covers the engineering insights required to frame the 2.4B parameter transformer, and eliminate audio artifacts under hardware constraints.

Read Story
Experiments |

What I Did Over My Summer Vacation

A summer vacation on an Alaskan cruise turned into a small fleet of side projects: a live ship tracker for the family back home, a PDF-to-EPUB pipeline so I could read arXiv papers on a Kobo, an EPUB builder for a papal encyclical, and a batch of updates to my Elvish linguistics work. Here is the roundup.

Read Story
Engineering |

Shipping the A2A Stack: A One-Day Build Log

In one day, five projects moved the Agent2Agent stack from demo to provable system: a dual-protocol lexicon server got durable cross-instance task state, the CLI shipped four releases with OAuth 2.1 and multimodal I/O, the reference services pinned down conformance, a TTS tool became an embeddable A2A skill, and the Flutter client got a byte-provable A2UI v1.0 target.

Read Story
Experiments |

Sandboxing Browser Agents — Isolation Options for Go + Chrome

There's a line in every Go browser agent codebase that disables Chrome's internal sandbox. Here's what it costs, and four concrete options for putting protection back — covering Apple Container, Docker, Cloud Run Jobs, and GKE Agent Sandbox.

Read Story
Experiments |

Computer Use Grows Up — Migrating to Gemini 3.5 Flash

We built a Go agent harness on Gemini 2.5's computer use preview. When 3.5 Flash landed, every action name broke on the first run. Here is the complete map of what changed, the register-both-branch-nothing strategy that kept it one codebase, and the validation-side guard that cost us the most time.

Read Story
Experiments |

Porting MisoTTS 8B to Apple Silicon: Inner-Loop JIT Compilation, 4-Bit Quantization, and Autoregressive Drift Mitigation

Ported the 8.2-billion-parameter MisoTTS model to Apple's MLX framework, unlocking local, real-time speech generation on macOS with a 6.7x speedup. By leveraging inner-decoder JIT compilation and 4-bit quantization, the implementation successfully bypasses Apple Silicon memory bandwidth constraints while resolving complex cross-platform numerical precision bottlenecks.

Read Story
Projects |

Interactive Phonology & Metrical Stress Analysis Live on Mithlond Linguistics

The Mithlond Project models Tolkien's Elvish as a dynamic historical language family rather than a static translation cipher. It features a custom phonotactic engine and interactive dashboards to programmatically analyze and visualize complex phonetic mutations, metrical stress shifts, and evolving structural sound laws.

Read Story
Engineering |

Taming Gemma 4 QAT: Loader Fixes, MLOps Journals, and Agent Skills on Apple Silicon

Local model fine-tuning on Apple Silicon is productive—until you collide with upstream parser bugs or silent weight loading issues. Here is how we tamed Gemma 4 QAT, established experiment journals, and created agent skills.

Read Story
Projects |

Mithlond: a return to my linguistic roots

The Mithlond Project brings J.R.R. Tolkien's invented languages to life through real-time neural speech synthesis, on-device mobile LLMs, and multimodal semantic search. By bridging rich linguistic lore with cutting-edge engineering, it transforms silent dictionaries into a fully audible and interactive experience.

Read Story
Experiments |

Bridging the Day 1 Tooling Gap: Fine-Tuning Multimodal Gemma 4 on Apple Silicon with mlxtune

I built an Apple Silicon-optimized MLOps toolkit to overcome the "Day 1 Tooling Gap" for cutting-edge multimodal AI models like Gemma 4. Featuring an end-to-end fine-tuning pipeline and a native Swift monitoring application, it empowers developers to train, visualize, and deploy complex audio-capable models entirely on a Mac.

Read Story
Engineering |

Giving Gemma 4 Ears: Porting Multimodal Audio to Apple Silicon with MLX Swift

Running Gemma 4’s audio pipeline natively on Apple Silicon requires moving beyond Python. I ported the Gemma 4 Audio pipeline from Python into pure, native Swift.

Read Story
Engineering |

Bringing Kokoro TTS to the Command Line: Building an Agent-Ready CLI with Rust

Fast, local, and high-quality Text-to-Speech (TTS) is becoming a necessity for developers and AI agents alike. Here is my journey building an agent-ready CLI for Kokoro v1.0.

Read Story
Experiments |

Multimodal Ground Truth: Building AudioVoxBench

I built AudioVoxBench, a native Swift benchmarking suite, to evaluate multimodal embedding strategies for an AI-generated music application. The data revealed a counterintuitive truth: rich text-augmentation actually outperforms raw media interleaving, achieving perfect semantic search precision at a fraction of the computing cost.

Read Story
Engineering |

Building apikeyscan: Triage for the Gemini Privilege Escalation

When Google's Gemini API silently turned historically public API keys into sensitive credentials, it exposed a new attack surface. I built `apikeyscan`, a Go-based CLI tool to rapidly audit Google Cloud projects and flag vulnerable, unrestricted keys before they can be exploited.

Read Story
Engineering |

Building MorphLab: Automating SVG Animations with Gemini 3.1

Why this matters: Animated SVGs are infinitely scalable and orders of magnitude smaller than GIFs or WebMs, but they historically required tedious hand-coding of SMIL tags. Gemini 3.1 Pro changes the equation, turning the "dark art" of vector math into a conversational interface.

Read Story
Experiments |

Podcasting the OpenClaw Zeitgeist: The Moltbook Trashfire

I asked an AI agent to make a podcast about a machine-generated social network. Instead, it built a defensive airlock, a Go-based recursive fetcher, and a counter-intelligence unit to survive the noise. A dispatch from the Moltbook trenches.

Read Story
Engineering |

Strike Price: Building a Jumbo Mortgage Strategist with Go and Gemini Live

Why this matters: Managing a jumbo refinance isn't just about watching a rate; it's about reserve requirements, seasoning rules, and the 10-Year Yield. Here is how I built an interactive command center to hunt for the perfect window.

Read Story
AI & Hardware |

Mattey Banana: Bridging Generative AI and CNC with Rust & WASM

How I built a web-based pipeline to convert AI-generated imagery into optimized vector paths for legacy CNC hardware using Rust, WASM, and Gemini.

Read Story
Engineering |

Building Steamer: A Resilient DNS TUI with Go

DNS management shouldn't be a chore. Steamer is an interactive TUI built with Go and Bubble Tea to make managing Porkbun domains fast and resilient. Here is what I learned building it.

Read Story
Experiments |

Pandemonium: Scaling Generative Social Simulation with Go

Pandemonium is a high-performance Go implementation of DeepMind’s Concordia, engineered to scale generative agent-based modeling for massive, distributed social simulations. By integrating native concurrency and the A2A protocol, it provides a robust, high-fidelity foundation for the next generation of autonomous digital societies.

Read Story
Thought experiments |

Agent Experience (AX) - Sofware for Agents

The North Star of development is shifting from UX to Agentic Experience (AX), where software survival depends on being discoverable and token-efficient for AI agents. This entry explores the rise of Inference Experience Optimization and how to prevent your code from becoming "dark matter" in an agent-led economy.

Read Story
Experiments |

Designing A2A Services for Long-Running Agents: Architecting for Patience

Moving beyond the "stateless wrapper" model, this project explores an architectural shift that moves AI state management from the client to the infrastructure. By combining the Gemini Interactions API with the A2A Protocol, I’ve developed a pattern for resilient, long-running "thinking" states and zero-copy handoffs between distributed agents.

Read Story
Experiments |

That Pesky WASM: Debugging WASM Crashes: When Parallelism and GIFs Collide

Gemini CLI resolved critical architectural bottlenecks and memory safety issues in Pixie-Anim by implementing a compositing virtual canvas in Rust to handle complex WebAssembly frame-buffer logic. This optimization ensures a stable, high-performance bridge between Rust and JavaScript for seamless GIF processing across both CLI and web environments.

Read Story
Experiments |

Things I did over the Winter Break ❄️

A winter break deep dive into refining high-performance tools, featuring Rust-powered GIF optimization and AI-driven idea expansion. These updates bridge the gap between creative brainstorming and functional execution across graph UI design, agent-led analysis, and automation.

Read Story
Experiments |

Building a Subjective GIF Optimizer from Scratch with Rust and Gemini

Pixie-Anim is a high-performance Rust engine that redefines GIF optimization by using Gemini 3 Flash as a subjective judge of visual quality. By moving beyond traditional pixel math, this WASM-powered tool achieves superior compression through a novel AI-guided feedback loop that outperforms industry-standard benchmarks.

Read Story
Engineering |

Improving Docstats: A Journey in Quality Hill Climbing

Docstats is a multi-protocol server designed to provide scalable readability metrics for AI agents and developers across web, PDF, and cloud sources. This project demonstrates high-assurance software engineering through modular refactoring and the implementation of a data-driven "Golden Set" for rigorous performance benchmarking.

Read Story
Experiments |

Home Control with Gemini CLI

I leveraged Gemini CLI to build a custom Go-based command-line interface that discovers and controls Lutron lights and Sonos systems. This project showcases how LLMs can automate complex network protocols and device pairing to create a unified, developer-friendly home automation tool.

Read Story
Engineering |

LitFlow: Bringing Graph Orchestration to Web Components

LitFlow is a framework-agnostic graph orchestration engine built with Lit and Signals to bring high-performance, node-based UIs to any web environment. By leveraging the xyflow core, it provides a lightweight, Material 3-styled solution for building complex, reactive architectural visualizations without framework lock-in. https://litflow.dev

Read Story
Tooling |

Bridges between Graphviz and XYFlow

Developing a high-performance Go translator to turn static architecture diagrams into interactive web components.

Read Story
Engineering |

Building an AI Agent that Browses the Web

A deep dive into creating a browser automation agent using Google's Gemini 2.5 Computer Use model and Chromedp in Go.

Read Story
Engineering |

Building an AI Photographer 📸

Build a tool that acts as a "Virtual Photographer" using Go, Chromedp, and Gemini 2.0 Flash (Vision).

Read Story
Experiments |

Things I did over the Thanksgiving Break 🦃

An exploration of Gemini CLI, coding conventions, and building 4 apps in a week including Banana Weather and Perfect Nature.

Read Story