Neural operating system
Atelier OS
A multi-agent neural operating system — the orchestration layer that coordinates specialized agents into one working environment. Architected and shipped in a 52-hour sprint.
Read the case study →Robert McCullock · AI Systems Architect · Boston
I build autonomous AI systems — and run them in production.
I am Robert McCullock, founder and architect of Design Delight Studio in Boston. I design the constraints and acceptance criteria; AI agents write, test, and deploy the code. Everything below is live, audited, and verifiable — led by the 12-agent Sovereign Orchestrator Pro, which publishes across four platforms on its own for under $13 a day.
Robert McCullock is a Boston-based AI systems architect and full-stack engineer, and the founder of Design Delight Studio. Working solo, he designs constraints and directs AI agents — Claude, Antigravity, local Ollama models, and Gemini — to build and ship production software. His work includes six production AI systems led by the 12-agent Sovereign Orchestrator Pro, two commercial desktop apps, two browser games, a live AI-built web app, a 59-class coding academy, and a 684-product Shopify store, all on local infrastructure at $0/month hosting.
110Store pages
51,000+Lines of code
$0/mo hosting
Contact: Robert@ddsboston.com · (617) 334-5912 · Boston, MA · linkedin.com/in/robert-mccullock-8580a8140 · ddsboston.com
The method
How I build: constraints, not keystrokes
I run a development pipeline where I write the spec and the acceptance criteria, and AI agents do the typing. Claude (Cowork) authors the code and a validation-gated deploy contract; Antigravity executes it through the Shopify MCP server; Cowork audits with SHA-256 and marker checks and a live smoke test; then I publish, with automatic rollback on any mismatch. No human keystrokes enter the production pipeline.
01
Architect
intent & constraints
02
Cowork
deploy contract
03
Antigravity
execute via Shopify MCP
04
Audit
sha256 · markers · smoke
05
Publish
human gate + rollback
rollback to previous on any mismatch
“I didn’t write a line of this. I designed the constraints. The agents did the rest.”Robert McCullock — Founder & Architect
# build manifest — this résumé page
template ......... page.portfolio.liquid
method ........... constraint-driven agentic build
engine ........... native scroll-driven CSS + worker canvas, zero frameworks
human_keystrokes_in_body ... 0
links ............ 18 internal targets verified live via Shopify Admin API (2026-06-13)
numbers .......... traced to forensic manifest (2026-06-06)
The systems
Six AI systems in production
Six working, audited multi-agent systems — more than 51,000 lines of code across the instrumented builds alone. One runs an autonomous publishing fleet, one is a neural operating system, one an autonomous digital office, one a generative-media agency, one a niche engine, and one a live customer-facing try-on. Counts are read from the June 6, 2026 forensic manifest.
Flagship system · autonomous · in production
Sovereign Orchestrator Pro
A 12-agent content engine that runs itself
SOP researches, writes, and publishes across four platforms — WordPress, Shopify, X, and Reddit — with no human in the loop. It runs on a hybrid stack: five local Ollama models on an RTX 3060 for the bulk work, with Gemini for the hard calls, so cost is telemetry-capped at roughly $8–$13 per day and often runs under $1. Seventeen concurrent threads, twelve specialized agents, JSON-persisted state, and a cost governor that throttles itself before it ever overspends.
And five more systems in production:
Autonomous digital office
AGI Nexus v9.0
An autonomous digital-office system that coordinates research, planning, and execution across a roster of agents — a technical-whiteboard build now in its ninth major version.
Read the case study →Autonomous creative agency
The Synthetic Director
A full-stack generative-media system (AGI-CORE-Pro V1.0) that scripts, generates, and assembles video and imagery through Veo 3.1, Imagen 4.0, and Gemini. The economics are the point: roughly $0.40 per 10-second clip versus $5,000–$15,000 for comparable studio production.
Read the case study →Niche & product engine
NICHE-FORGE-CORE V3.0
A research-to-output engine built on a dual three-step pipeline across four Gemini models, turning a niche brief into structured, ready-to-use product and content assets.
Read the case study →Live · customer-facing
Shopify Virtual Try-On
Photorealistic AI garment fitting on apparel product pages, powered by gemini-2.5-flash-image. A key-safe Express proxy on GCP Cloud Run keeps the API key server-side. No email, no signup — verified live on the published theme.
Open the Try-On Studio →Read deeper
Full case studies
Each system has a dedicated technical write-up, and the capability brief frames the whole portfolio.
Capability & traction brief →Live & purchasable
Two commercial desktop apps, on sale now
Not prototypes. Two Electron desktop apps ship today on ddsboston.com, both bring-your-own-key with a serverless license gate, so inference cost sits with the user and there is no subscription. VibeTube was built in 21 hours.
Generative audio · desktop
PromptDJ NS-9000
$79.00 USD
A bring-your-own-key neural synthesizer — an Electron desktop client that turns prompts into generative audio, gated by a serverless license check.
Buy PromptDJ →YouTube automation · desktop
VibeTube AI Desktop App
$49.99 USD
A bring-your-own-key YouTube content-automation studio in a single Electron desktop app — a 19-section orchestration matrix turning a URL into titles, tags, scripts, thumbnails, and cross-posts, with the same serverless license model and zero subscription.
Buy VibeTube →Free & live
Live builds, games, and tools
Beyond the products: a live AI-built web app, two free browser games, and two free tools — all shipped and in use, all built with the same AI-directed workflow.
Live web app · AI-built
The Bay State Treasure Gazette
A real-time directory of Massachusetts thrift, vintage, and antique shops — a single-page web app with structured data and an llms.txt, engineered to be found by both search engines and AI assistants.
Visit the Gazette →Free browser game
Legend of the Red Dragon 2026
A modern, AI-powered reimagining of the BBS-era RPG classic — turn-based combat, character progression, and a living text world, rebuilt for the browser.
Play LORD 2026 →Free browser game
Dragonfly Crush
A polished match-3 puzzle game that respects your privacy — no email, no signup, no tracking. Pure browser gameplay, built end to end.
Play Dragonfly Crush →For Shopify owners
One-Prompt App Library
A library of single-prompt app builds for Shopify store owners — copy a prompt, get a working tool. A teaching artifact and a working toolkit in one.
Open the library →For tenants
Free Legal Help — Housing Defense
A free, self-serve tool built to help people defend their homes — a real civic build, live and in use, made with the same AI-directed method.
See the tool →Old-internet ethos
Free, no ads, no pitch
The games and tools are free and unmonetized — 1990s-BBS nostalgia rebuilt with current tooling, shared rather than sold.
Hard-won lessons
Engineering judgment, from real incidents
The work that does not fit in a feature list: production failures diagnosed and fixed. Each card is a real incident from a build session — symptom, root cause, fix, and the lesson it locked in. This is the difference between writing code and operating systems.
Symptom
Cleaning ~6,000 legacy spam blog posts without wrecking SEO equity.
Root cause
A retired legacy AI system had auto-published 8,254 articles; most were disposable, a few had value.
Fix
HTTP 410 (Gone) for the bulk — faster de-index than 404, equity deliberately forfeited — with 301s only for the top 100, behind backup JSON and pre-flight abort gates. 6,163 deleted, 100 redirected, 1,995 kept, in 1:38:01, zero errors.
Lesson
Destructive mass operations need restore points and tolerance checks before they run.
Symptom
Storefront product filters returned zero results for every category.
Root cause
Print-on-demand set one product_type — "Print Material" — on every product, so type-based filtering had nothing to split on.
Fix
Moved category detection to the product title, and gated the "Certified" badge to apparel only.
Lesson
Verify the shape of the data the platform actually emits before trusting a field.
Symptom
A Gemini model upgrade silently killed The Synthetic Director's posting workflows.
Root cause
Model IDs were hardcoded across 22 files; a duplicate <script> tag was double-mounting React; a 35KB brand-bible file was orphaned from the runtime.
Fix
Centralized the model IDs, removed the duplicate mount, wired the brand file back in.
Lesson
Hardcoded vendor identifiers are a time bomb; config belongs in one place.
Symptom
Map tooltips on the Academy constellation hid behind neighboring nodes.
Root cause
transform: translate on each node created a new stacking context, trapping the tooltip; flip-left and flip-right class bodies were also semantically swapped, and translucent backgrounds let the gold core bleed through the text.
Fix
Switched to negative-margin offsets and two targeted str_replace patches — no full rebuild.
Lesson
Know which CSS properties create stacking contexts before you reach for them.
Symptom
Mobile performance on Shopify Basic plateaued no matter what was tried.
Root cause
Platform-injected scripts set a hard floor on Total Blocking Time.
Fix
Pushed mobile Performance 62 → 76 and TBT 160ms → 20ms, then proved ~78–82 is the realistic ceiling.
Lesson
Knowing where the ceiling is stops you wasting effort chasing a 90 that the plan can't reach.
Symptom
A multi-model local matrix on one consumer GPU thrashed and stalled.
Root cause
Concurrent model calls oversubscribed the 12GB VRAM on an RTX 3060.
Fix
Async-mutex serialization of model calls and a num_ctx budget of 4096 — production inference on hardware already owned, at $0/month hosting.
Lesson
Treat VRAM as a shared resource with a scheduler, not an infinite pool.
Symptom
Duplicate JSON-LD kept appearing across theme sections.
Root cause
Multiple sections each emitted their own Organization / ItemList schema.
Fix
One schema, one owner: theme.liquid owns Organization/WebSite/Breadcrumb; the hero owns ClothingStore/FAQPage; the SEO manager owns ItemList/HowTo. Collisions eliminated.
Lesson
Structured data needs a single source of truth per type, like any other state.
Symptom
An agent skill could hallucinate brand facts or ship code that broke house rules.
Root cause
The agents had no encoded ground truth for the brand or the platform's limits.
Fix
Wrote a 5,529-LOC brand-rules skill, then validated it with a blind build test — a fresh agent passed all 10 proof-required self-audit items on the first attempt.
Lesson
Prove a skill against a cold agent; don't trust it because you wrote it.
How I work
Operating principles
“No marketing fluff.”
Only real, verifiable things. Claims are sourced or they don't ship.
“Accuracy beats completeness.”
When a figure is uncertain it gets marked, checked, or cut — not rounded up.
Proof over vibes.
Every build is audited to 100% — marker counts, SHA-256, live smoke tests — before it goes live.
A worked example of that discipline: a June 2026 forensic pass over my own codebases corrected my published figures in both directions — the Synthetic Director's persistence was relabeled to Dexie/IndexedDB, a font claim was corrected, and the academy count moved up. The codebase won.
Always on
An autonomous operations fleet
Beyond the headline systems, eight or more scheduled agents run the day-to-day with the human out of the loop — research, publishing, video, and engagement — reporting in rather than waiting for instructions.
Daily
Morning Brief
An autonomous research-to-publish agent that has shipped 25+ consecutive daily blog issues — each with JSON-LD, a CDN cover image, and FAQ structure — without manual authoring.
Daily
Video build
A scheduled pipeline that drives a browser video editor end to end — builds, exports, and downloads finished shorts — every day.
Continuous
YouTube engagement
Agents that verify scheduled videos went public, then post and pin formatted comments — and yield safely when another agent is writing.
Weekly
Long-form drafters
Recurring agents that draft long-form threads and articles on a weekly cadence for human review.
Teaching the method
The DDS Vibe Academy — 59 free classes
I teach the method openly. The Academy is a free, 59-class curriculum on intent-based AI coding, delivered across three gateway templates and built with the very workflow it documents — its public build manifest records zero human keystrokes in the build pipeline. Nothing is paywalled.
Curriculum
59 classes
From "what is vibe coding" foundations through agentic builds, MCP servers, sovereign RAG, and the SEO Magnet System.
Access
Free, no gate
The entire curriculum is openly accessible — built as a teaching asset, not a paywall.
Built by AI
3 gateway templates
Delivered across three template variants, every page assembled with the constraint-driven workflow on this page.
The storefront
A 684-product store, engineered solo
Design Delight Studio is itself a proof of work: a full sustainable-apparel catalog on Shopify Basic, built and tuned by one architect. Performance engineering cut largest-contentful-paint from about 12.5 seconds to 3–5 seconds and lifted PageSpeed from 47 into the 70–80 range.
Catalog
684 products
Made-to-order sustainable apparel and accessories across 110 pages — tees from $24.99, totes $34.99.
Accessibility
WCAG 2.1 AA
20+ store sections rebuilt to meet WCAG 2.1 AA, with semantic structure and keyboard support.
Core Web Vitals
LCP 12.5s → 3–5s
PageSpeed lifted from 47 into the 70–80 range through deferred loading and image and font discipline.
Discovery
7 JSON-LD / page
The SEO Magnet System: structured data, answer capsules, and a 2,388-line sitemap engineered for AI citation.
Home
3,178-line section
A single hand-architected homepage section coordinating the store's primary merchandising and narrative.
Reference
The method, documented
The SEO Magnet System is published openly as a masterclass.
SEO Magnet masterclass →The stack
Technical skills
The tools behind the work — from multi-agent orchestration and local LLM inference to production Shopify commerce and Core Web Vitals tuning.
AI & agents
Multi-agent systems
Frontend
React & TypeScript
Backend & data
Node, Python, cloud
Commerce & ops
Shopify & tooling
Experience
Founder & CEO — Design Delight Studio
Boston, MA · 2024–Present · AI Systems Architect
- Designed and operate an AI development pipeline in which agents write, audit, and deploy production code under written constraints — zero human keystrokes in the production pipeline.
- Shipped six production AI systems (250+ source files, 51,000+ LOC across instrumented builds), two commercial desktop apps, two browser games, and a live AI-built web app — solo.
- Built and operate a fleet of 8+ scheduled autonomous agents handling daily publishing, video, and engagement.
- Built and run a 684-product Shopify Basic storefront; cut LCP ~12.5s→3–5s and lifted PageSpeed 47→70–80 across 20+ WCAG 2.1 AA section rebuilds.
- Created the free 59-class DDS Vibe Academy and a 5,529-LOC agent skill that encodes brand and platform rules, validated by blind build test.
- Manage five supply-chain certifications (GOTS, GRS, OCS, PETA-Approved Vegan, Fair Trade) across a made-to-order catalog.
Reach
Audience built from zero, with $0 ad spend
Distribution is organic across seven platforms with no paid advertising. The figures below are a June 2026 snapshot.
46,875YouTube lifetime views
115YouTube subscribers
+250%Subs, last 28 days
2,321X followers (@ddsboston24)
16,500+X posts
$0Paid ad spend
Two of the systems are commercial products on sale today, and the storefront has been live throughout. Specific sales figures are kept private; the proof on this page is the shipped work itself.
The brand
Five sustainability certifications
Design Delight Studio's apparel is made to order against recognized standards.
GOTS
GRS
OCS
PETA-Approved Vegan
Fair Trade
Design Delight Studio is not OEKO-TEX certified. See the certifications page →
Questions
FAQ
Who is Robert McCullock?
Robert McCullock is a Boston-based AI systems architect and full-stack engineer, and the founder of Design Delight Studio. Working solo, he designs constraints and directs AI agents to build, test, and ship production software.
What does Robert build?
Autonomous, multi-agent AI systems and the products around them: six production systems led by the 12-agent Sovereign Orchestrator Pro, two commercial desktop apps, two browser games, a live AI-built web app, a 59-class coding academy, and a 684-product Shopify store.
How does one person ship this much software?
Through a constraint-driven workflow. Robert writes the spec and acceptance criteria; Claude (Cowork) authors code and a validation-gated deploy contract; Antigravity executes it through the Shopify MCP server; Cowork audits with SHA-256 and marker checks and a live smoke test. No human keystrokes enter the production pipeline.
What is Robert's technical stack?
React 19, TypeScript, and Electron on the frontend; Node.js, Python, Firebase, GCP Cloud Run, and IndexedDB on the backend; Shopify Liquid, 2.0 sections, and JSON-LD on commerce; and multi-agent orchestration with local Ollama models, Gemini, and Claude, including MCP server development.
Are the systems actually in production?
Yes. Sovereign Orchestrator Pro publishes across four platforms daily; the Virtual Try-On runs live on apparel product pages; the Synthetic Director, NICHE-FORGE-CORE, Atelier OS, and AGI Nexus are working audited builds; and a fleet of scheduled agents publishes a blog issue, produces video, and runs engagement every day.
What are the two commercial apps?
PromptDJ NS-9000 ($79) and VibeTube AI Desktop App ($49.99), both sold on ddsboston.com. They are bring-your-own-key Electron apps with a serverless license gate, so inference cost stays with the user.
What is the DDS Vibe Academy?
A free, 59-class AI coding curriculum on ddsboston.com, built end to end with the same AI-directed workflow. Its public build manifest documents zero human keystrokes in the build pipeline.
How do I work with Robert?
Email Robert@ddsboston.com. The case studies, the desktop-apps portfolio, and the capability brief document the work in depth.
In one screen
Key takeaways
- 6 AI systems in production, led by Sovereign Orchestrator Pro (12 agents, 17 threads, 4 platforms), plus 2 desktop apps, 2 games, and a live AI-built web app.
- 250+ source files, 51,000+ lines of code across the instrumented systems — counted from a forensic manifest.
- 684 products and 110 pages, built and run solo on Shopify Basic; LCP ~12.5s → 3–5s, PageSpeed 47 → 70–80.
- 59 free classes in the DDS Vibe Academy; a 5,529-LOC agent skill validated by blind build test.
- $0/month inference on local models; $0 paid ad spend across 7 platforms.
- Zero human keystrokes in the production pipeline — the architect writes constraints; agents write the code.
One architect, directing AI agents under tight constraints, shipping and operating production software at the scale of a small company. The apps sell, the store runs, the academy teaches, and the systems publish every day — all verifiable from this page. If that is the kind of builder you need, let's talk.
Work with Robert
Let’s work together
I am open to senior AI-engineering and architecture roles, consulting, and partnerships. The fastest path is a direct email — or dig into the proof first.
Boston, MA · (617) 334-5912 · Robert@ddsboston.com