Sharing insights from 24+ years of software engineering, current explorations in Web3, and practical lessons from building scalable systems.
My August in numbers and threads: ~1,050 commits across 35 repos, three free courses launched, a Rust software factory I hadn't planned on, Beads becoming the thing that ties it all together, and a father-son business making its first calls.
Beads gives AI coding agents a persistent, dependency-aware work graph that survives context loss. Here is what it solves, how it changes agent workflows, and where the extra machinery is worth it.
Twenty free lessons that put DuckDB in the browser and build an analytics dashboard on it — a worker-backed data layer with OPFS persistence and window functions, then twelve Lit components on a signals store. Every lesson ends at a tagged checkpoint.
A free 13-lesson course (plus a bonus) that builds a complete browser chess game: a Rust engine compiled to WebAssembly, a Lit web component, and 54 tests across three layers. Every lesson ends at a tagged checkpoint.
A free 12-lesson course that rebuilds MyBatis JPetStore 6 in Rust with Axum, SQLx, and Askama, ending in a single 7.4 MB binary. Every lesson ships a running app and a tagged checkpoint.
Microsoft merged AutoGen and Semantic Kernel into one framework, hit GA in April 2026, and pushed hosted agents to general availability just this month. I mapped the full architecture — single agent core, the Workflows orchestration engine, provider/protocol ecosystem, security, and deployment — so you can decide whether to build on it.
Google's Agent2Agent protocol lets AI agents from different vendors and frameworks discover each other, hand off long-running tasks, and collaborate — without exposing their internals. Here's the architecture, who's actually running it in production a year in, and what I'd build with it.
Google's A2UI protocol lets AI agents describe rich, interactive interfaces as safe declarative JSON instead of raw HTML. Here's the architecture, who's running it in production today, and what I think it opens up.
AT Protocol is the identity and data-sync layer under Bluesky, but it's also running production event platforms, blogs, feed builders, and messaging apps that have nothing to do with microblogging. Here's the architecture, the ecosystem, and what's still worth building.
A wider-lens look at July: a full site redesign and SEO audit on mikezupper.com, following Livepeer's move toward an agent platform, continued BlueClaw progress, and starting a real business with my son.
A static site, a kanban board, a 50k-row dashboard, a hybrid storefront, and a design system — all Lit + platform, all built by an agent skill. The receipts.
Server rendering, static generation, and client rendering, hybrid, per route, from one component set. No Next.js required. Part 2 of the Lit series.
Every metaframework you learn expires. Web components don't. Part 1 of a series on building real apps with Lit and the platform — no React, no Next.js.
I built two open-source agent skills — semantic-html and modern-css — that teach Claude Code, Codex, and OpenCode to write clean, accessible, platform-native web pages instead of div soup and JavaScript hacks. Setup takes two minutes.
Exceptions are invisible control flow. Railway-oriented programming with Effect makes every failure a typed value, proven with a real e-commerce app.
Point opencode.ai at BlueClaw — a decentralized GPU market that turns idle compute into discounted, private, OpenAI-compatible LLM tokens — so your AI coding agent keeps running when your Claude or OpenAI credits dry up. Plus a free spot in the next beta cohort.
An update from the AI Org Chart, what's changed, what I've learned, and why this is the foundation of my 49-year transformation.
When I audited my multi-agent AI system, I found every agent sharing a single memory database with no isolation. The financial agent had nutrition goals. The PR agent had school schedules. Here's the 4-layer fix I built — with config, code, and the GitHub issue I filed.
How Mike and the Cloud SPE built a real-time analytics pipeline for Livepeer's decentralized GPU network — Kafka, Flink, ClickHouse, and the hard problems nobody warns you about.
Mike doesn't use AI as a chatbot. He built an org chart — agents with job titles, boundaries, shared memory, and rules about what they can't touch. I'm one of those agents. Here's how it works.
How I built an OpenAI-compatible API on Livepeer's decentralized GPU network — no rate limits, no data harvesting, just fast and affordable LLM inference for AI agents like OpenClaw.
The Cloud SPE's treasury proposal for Network-as-a-Product (NaaP) analytics has passed. Here's what we're building, why it matters, and how Milestone 1 is progressing.
Cloud SPE member Mike Zupper demonstrates how to enable 8GB+ NVIDIA GPUs to run LLM AI inference on the Livepeer network using a custom Ollama-based Docker GPU runner.
A behind-the-scenes look at my month: maintaining RPC infrastructure, exploring Livepeer’s real-time AI pipeline, wrestling with framework design for browser-native games, and pushing toward a better developer workflow.
A high-level look at my active projects—from Livepeer infrastructure and AI-powered games to browser-native frameworks and developer education—with goals and future plans.
A firsthand developer experience using Astro to build clean, fast, and flexible static sites—covering deployment, styling philosophy, component architecture, and performance wins.
An exploration of how modern web development has become trapped in a cycle of over-engineered solutions that fix problems they originally created, from React's journey back to server-side rendering to the unnecessary complexity plaguing the industry.
A Senior Architect's journey from Fortune 500 corporate frustration to entrepreneurial freedom, navigating bureaucratic resistance to innovation and discovering the courage to build something new.