These case studies are about the decisions most people never see: the rules that keep products coherent, behavior predictable, and teams moving in the same direction. Each one examines a different layer of that work—from design systems and platform UX to multimodal interaction and AI behavior—and asks the same question: will it still hold up when the product, the organization, and the stakes get bigger?
Each of these looks at the structure underneath a product rather than its features: how several surfaces manage to coexist, how behavior stays predictable as things grow, and whether the decisions still hold up a few years later. Every case study takes a different layer of that problem.
A design system only scales if quality is the default — not something every team has to re-earn on each screen.
A Home for All Our Stuff Google · design-system hubOne governed home for an organization's design knowledge — inheritance, inline editing, a publish gate, and a constellation of connected tools.
Window Management System Platform UXWindows only behave when the rules live above the apps, so nothing has to trust anything else to get it right.
Widget Drawer Ambient surfacesInformation that reaches you without costing you attention, governed from above the apps rather than inside them.
Mobility AI In-vehicle AIAn assistant in a moving car has to earn the interruption — which mostly means knowing when to wait.
Behavioral Orchestration Behavioral frameworkHow should software behave when a wrong move costs something real? A framework for when to speak, wait, escalate, or stay quiet.
One question, tested across very different projects: when should a system act, when should it ask, and when should it hold? Each one is built to show its work — every claim traceable back to a source, and a system that would rather decline than guess. Built solo, with AI, in 2026. A couple more stay private.
An ambient practice-minder for solo attorneys. It reads the case files they already have and keeps an honest picture of the practice — deadlines with provenance, contradictions with receipts, nothing unconfirmed ever counted as knowledge. v0.79, ~530 tests, local-first.
Component Studio Design-systems platform · foundation builtA career of design-systems work, turned into a product. Every component becomes a Passport — one canonical record that design tools and codebases consume instead of owning. 61 components, one pipeline, pixel-proven; 450 tests, 45 ADRs.
Atlas of American History Interactive history · in polishA source-driven atlas where presidencies anchor the time spine and every claim shows its receipts. 469 nodes, 838 connections, seven analytical lenses. Built by directing a fleet of AI agents through written briefs.
DiamondCoach Coaching tool · phase one shippedA lineup planner for youth softball and baseball coaches — the magnetic board, rebuilt for the iPad in the dugout. Fully offline, no accounts, no cloud. Made because I coach.
Treasures of Iran Cultural platform · growingPersian literary and material culture, made explorable — a Hafez divan of 495 ghazals, a 3D library timeline, and a design line that puts Rumi and Ferdowsi on shirts.
Behind the Mask Game · shipped web buildA card game about the faces people wear, co-founded with my wife — a 60-archetype framework, a 500-scenario dataset, and a playable browser build from an AI-native design-to-code loop.
Six products, one process. Hover any step — each one carries the stack it runs on and a receipt from the lab.
Hover a step to see its stack and the receipt behind it.
Claude Code / the Hub · restraint-first, human-in-the-loop.
Problem space Solution space ProductionFrame the problem, map the domain, benchmark what exists.
Human gateframe the problem
StackClaude · ChatGPT · primary-source research · synthetic corpora
ReceiptFez-Demo — synthetic attorney-grade corpus, 12 build sprints of domain immersion
IA, user flows, and a scope-lock. Every build starts as a written spec.
Human gatelock the scope
StackFigJam · Claude · written agent briefs — the spec
ReceiptAtlas — 15+ written agent briefs; 469 sourced nodes built from them
Hi-fi design on a token-based system; AI builds against the locked spec.
Human gateapprove the design
StackFigma · Figma MCP · Claude Code · React + TypeScript · design tokens
ReceiptComponent Studio — 61 components, one render pipeline, 45 ADRs
Deterministic tests, adversarial evals, and human usability audits.
Human gateaccept the evals
Stackmy QA harness · pixel-parity goldens · evaluator protocols · human audits
ReceiptChambers — ~530 tests · 3-round evaluator protocol · attorney usability audit
Specs, handoff package, and the ship itself.
Human gateapprove the ship
StackFigma Dev Mode · git-tagged releases · the Hub
ReceiptShipped live on pej.me — Atlas · DiamondCoach · Behind the Mask
Run it live, watch behavior, feed learnings into the next cycle.
Human gatejudge the results
Stackthe Hub · release cadence · daily hands-on use
ReceiptChambers — ~125 tagged releases, local-first, in daily use