// Case studies
Build logs, with the wrong turns left in.
Architectural decisions written up while they were still reversible. Every number traces to a repository you can read.
anthill: An Agent Factory That Publishes Its Own Error Rate
An agent factory with a code reviewer inside it. A person types a request, a cloud model plans and writes the agent, and it lands in the repository under one property held by a test rather than a promise: merged but never scheduled, because merging is reversible and scheduling is not. The reviewer is one agent of nine, running on an 8B model on a Mac mini with Claude Opus auditing its work periodically. Over 714 hours it completed 8,826 review runs across 2480 versions of 1291 files in two repositories and pushed 24.88 million prompt tokens through the local model at no marginal cost. The paid audit has now graded 715 of its findings, upheld 12, and rejected 703. That result is the case study. Local inference solves cost and privacy; it does not solve quality, and running it honestly means measuring the gap rather than reporting the volume. What makes the design worth keeping is that the model is allowed to be wrong cheaply: a free, no-model grounding filter discards 572 findings that quote code which is not there, before a cent is spent. The uncomfortable update is that scale changed nothing: a month ago this read 3 of 182 upheld, a precision of 1.6%, and at four times the volume it reads 12 of 715, a precision of 1.7%. The verification stage responsible was diagnosed and rebuilt because it was refuting nothing at all, and it now refutes 37.5% of what it sees, yet the rate the paid layer upholds went down rather than up. This page reports that rather than presenting the intermediate number as a win. Next phase: giving each agent a speaking avatar driven by its own voice.
The G Training Center: A Null Baseline for an Agent Persona
A local harness lab for G, the assistant inside GBuild, built because a cloud build took most of an hour while the thing being changed was a text file. The lab points the installed app’s own engine at a swappable home directory, so a persona edit is live on the next run with no rebuild and no release. The part worth reading is the measurement design rather than the speed. GBuild carried 33 expertise overlays that had never been validated, and the honest first question was not which to migrate but whether the overlay changed any outcome at all. A stochastic engine answers the same prompt differently every time, so a raw difference between treatment and control can never justify deletion. The A/B therefore runs three arms and makes two of them identical: the gap between the two identical arms is the engine’s own noise floor, and an overlay has to clear it before it counts. Grading is deliberately left to a human, because an automated judge would be a second unmeasured layer scoring the first, which is the exact stack the lab exists to stop. An ungraded arm reports pending, never a fabricated pass.
The Agent Identity Stack: Giving an Agent Its Own Email, Number, Domain, and Card
By mid-2026 an autonomous agent can get a wallet off the shelf and a payment protocol off the shelf, but name and reachability (domain, DNS, mailbox, phone number) are still human-gated and priced for enterprise SaaS. Money got an API before identity did. This is the four-organ stack built to close that gap for GBuild’s agent: its own inbox on Cloudflare Email Routing, its own Telnyx number, its own domain via Latch, and a Stripe Issuing card whose budget is enforced by the card network rather than by our code. The safety property that shapes the whole design is that the agent cannot mint its own card or write its own DNS: both are human-plane only, pinned by tests. The card path is TEST mode and the proactive capability loop is deliberately inert; the case study renders per-organ status rather than claiming four shipped things.
Latch: DNS Automation an Agent Can Drive
Connecting a custom domain is the most common drop-off point in SaaS onboarding, and every product in that category assumes a human is sitting at the registrar login ready to click approve. An agent has no human at that step. Latch is the in-house DNS-automation service built to close that: detect where a domain’s DNS is hosted, write the records where it can, fall back to a guided manual flow where it cannot, and verify propagation server-side against multiple resolvers. The engineering that matters is the email-DNS intelligence: merging into a single SPF record instead of adding a second, refusing at the 10-lookup ceiling rather than silently flattening, and hard-stopping on any existing MX. Deployed at latch.gbuild.app; the client transport reaches devices in the next sidecar release.
The Foundry: An On-Device Model Line with Promotion Gates
GBuild’s own small models, held to a measurement discipline: synthetic training data generated by typed-field corruption (so it costs nothing), human-labeled gold sets, paired-bootstrap 95% LCB significance bars, and champion/challenger promotion. The first shipped model is a Rust runtime contact ranker with a versioned feature contract to its Python trainer. The keystone grounding verifier scored 97.5% on its own synthetic holdout and 62% on the gold set, a 35-point overfit the gates caught and the case study leads with. Currently building.
ThingStore: Supply-Chain Security for an App Store, Built Solo
A signed, sandboxed marketplace for AI-built apps inside GBuild. Every publish clears a 5-layer automated gate plus egress-allowlist scrutiny; every install verifies an Ed25519 publisher signature over a SHA-256 content hash before a byte runs; every installed app is confined to its own Electron session partition and a loopback egress broker that resolves DNS internally to block rebinding. Revocation holds offline. The known limits (a confirmed Seatbelt host-scoping defect, orchestrator-glue test gaps) are documented in a shipped inventory rather than hidden. Currently building.
Voice Agents: An Operating System You Can Talk To
The voice layer of GBuild: dictation, conversational, and realtime modes over OpenAI Realtime and Gemini Live (3.1 Flash Live), with on-device MLX Whisper STT and Kokoro TTS. An intent router turns speech into tool calls that operate macOS apps, files, and folders, GBuild itself, and the agent fleet, behind destructive-intent confirmation and transcript redaction. Includes a custom Voice Harness Lab where per-turn human feedback on real conversations becomes an approved patch to the harness source, a training center of scripted voice skits and router benchmarks, a proactive agent that completes a PRD and CRM records live from a running conversation, and phone-calling agents deployed internally in a test environment. Currently building.
PasteDrop: 8 Days From Idea to the Mac App Store
A macOS menu-bar clipboard utility shipped through a new ship-apps pipeline. The entire codebase was written in a 21-minute sitting across three commits, then landed on the Mac App Store eight days later. PasteDrop is the pilot of a repeatable scaffold + template + XcodeGen + CLAUDE.md + ASC playbook system.
Building PrayerMap: React Native + Supabase + Mapbox
Full-stack mobile app architecture: real-time prayer connections, Mapbox GL integration, 6-dashboard admin command center, and a 3-tier AI behavior monitoring engine.
Ethereal Glass: A Design System from My Idea to Figma to Production
How PrayerMap’s UI/UX was born in Figma as the "Ethereal Glass" design language, translated into a production React web app, then migrated to React Native with a 470-line animation preservation spec bridging Framer Motion to Reanimated and Skia.
Claude Code Toolkit: Production Infrastructure for AI Agents
An open-source toolkit adding agent orchestration, smart MCP loading, quality evaluation, session intelligence, and project bootstrapping to Claude Code.
Neurotoken: Adaptive Thinking Allocation for Claude Code
A prompt scoring engine that replaces static thinking allocation with a two-axis complexity × stakes matrix, 11 model/effort tiers, verb-aware parsing, context dampening, adversarial-tested signal tuning, and an opt-in ceiling mode for cost-optimized downgrade, all in zero-dependency Node.js.
Directors Chair: Mission Control for an Autonomous Agent Fleet
A real-time mission control dashboard monitoring an autonomous agent fleet across job search, project maintenance, and infrastructure.
Ora: An Orchestration Framework Extracted from Production
Extracted from the MCP-WP prototype. Coordinates Claude Code, Cursor, GitHub, Pinecone, LangSmith, and Slack into a unified development pipeline with file-based agent coordination and a 5-minute bootstrap, 15-minute daily oversight model.
MCP-WP: A LangGraph Architectural Prototype for WordPress Analysis
A LangGraph-based architectural prototype for multi-agent WordPress analysis. 27 agent personas built on 7 LangGraph agent types, with Pinecone vector retrieval and a live orchestration harness. Performance scoring and revision loops are stubbed; the sync layer and orchestration are production.
The n8n Advantage: 29 Workflows Automating Everything
How I built 29 production n8n workflows spanning AI code review, daily orchestration, Slack bots, Instagram content generation, and session intelligence classification.
LivCut: Privacy-First Video Editing (In Development)
Building a professional video editor with 66 Metal compute kernels, 7 on-device AI engines, 471 tests, and 257 Swift files. All processing runs locally on the Apple Neural Engine. In active development.
Puber: An Ultramarathon Support App, Built End to End
A community app for the 100-mile ultra world: connecting trail runners with pacers and crew, plus per-race coordination boards. Expo SDK 54 mobile app, a Next.js ops console, and a Supabase backend with RLS on every table, a Postgres match state machine, 14 edge functions, 5 pg_cron jobs, and an offline Race Kit. Safety was designed in as architecture. v1 was built end to end and ran locally. Development stopped before release, and it was never shipped to the App Store; this stays up as a record of the build, not as a product.
GBuild: An AI-Native Build System
A macOS desktop app that orchestrates a fleet of AI coding agents through 5 first-party drivers (Claude Code, OpenClaw, Hermes, Codex, shell + sandboxed bring-your-own) from a calm, voice-first interface. Electron + a Rust sidecar over JSON-RPC (1045 registered methods), pluggable agent drivers, an on-device voice pipeline (MLX Whisper, Kokoro, Gemini Live), a fully local Qwen3 RAG layer, bidirectional Obsidian sync, and full MCP lifecycle management. The resident agent behind the "G" persona is a fork of Nous Research’s open-source Hermes agent (MIT), vendored into the app with its provenance pinned at build time; GBuild deliberately stays the orchestration layer above it and does not adopt its work queue, so the engine remains one swappable driver among five rather than the system of record. One shell: the "Lattice" cockpit, plus a menu-bar ambient presence; the legacy power-user portal was retired in a 26,000-line single-shell consolidation. Behind it: an 18-Worker Cloudflare edge backend, a signed app marketplace, and an on-device model line. 344 notarized releases shipped in its first 121 days. Marketing site live at gbuild.app; the desktop app is in private development. Currently building.
Johnson Power Rankings: A World Ranking for the 200+ Mile Ultra
A data-driven, weekly world ranking system for ultramarathon runners, launching with the 200+ mile distance tier. A pure-TypeScript ranking engine (course-difficulty model, iterative field-strength, time-decayed weighted-mean ratings, backyard scoring) over a Supabase backend with 11 migrations, RLS on 20+ tables, 6 edge functions, and a Monday recompute cron. A Next.js web app, a separate admin panel, and an Expo mobile app. Built with my dad as the domain expert, in roughly two days. Live in preview; public launch pending. Currently building.