In June 2026 we audited 38 Danish SMB websites across five verticals through Citetome's automated pipeline, anonymized. The headline: average AI-visibility is 83/100 - assistants can read these sites just fine - but average agent-readiness is 48/100. Only 26 percent publish an llms.txt, and 0 of 38 publish structured prices, availability, or opening hours. Not one. An AI assistant can describe these businesses; an AI agent cannot transact with a single one of them. The gap is not a technology problem - Schema.org has carried prices and opening hours for over a decade - it is an intent problem. The full post covers what the two scores actually measure, why "readable" happened by accident while "transactable" requires a decision, and the one-day fix list we verified on a live client site this week: 81.7 to 100.0 AI-readability, 37 to 82 agent-readiness, zero design changes.
Redesigning the case-studies page as a Mission Control: 13 monitors, zero fake numbers
The old /cases/ page was a generic bento grid - card, card, card, card. Useful, navigable, forgettable. Today it became a Mission Control: one operator console with 13 monitor cards, each showing a custom widget that reflects what the project actually is. ElektraOS gets an agent grid pulsing 41 dots, one per agent. Loomgraph gets a node network with connecting lines that dash and pulse. Vienna Artists gets a waveform of 24 audio bars. AYLA Maison gets a gold shimmer typographic mark with floating particles. The Global Creative Network gets a stylised six-city map with corridor lines. MealShift gets a terminal scroll. The UK property brief gets an SVG floor plan of all eight units. The 40th Brick gets a CSS pixel grid spelling its number. Operations Suite gets a workflow diagram. Project N gets a CLASSIFIED glitch with redacted bars. Twelve distinct widgets, all CSS and SVG, no external JS, no images, no hallucinated activity. The Roster ticker at the top cycles through every project but uses only metrics already published on each card - no invented agent activity, no fake API logs, no "shipped 4 outbound drafts" filler. The hardest debugging moment: the global animations.css .reveal opacity:0 was hiding every monitor below the fold because the site-wide IntersectionObserver only fired for cards already in view at load. Override added at the page level; monitors render immediately, hover does the motion. The bigger lesson: when the work IS the differentiator, your portfolio should let people see into each project, not just past it.
EngineeringLaunch
From Loomgraph to Citetome: shipping a marketing site that runs on real verified data, not mockups
Loomgraph rebranded as Citetome and shipped a full marketing site in one day - hero, product card, pricing grid, audit request form, live dashboard. What made it different: every number on the page is wired to api.citetome.com, not invented. Six sites monitored, 69 entities catalogued, 141 tests run, 12 alerts open, 15 proposals pending. Per-site scores live on the page: elektraos.dev 0.92, sheandelle.com 0.89, aylamaison.com 0.83, mealshift.co.uk 0.77, freshmarket.ma 0.68, bizus.co.uk 0.57. No mockup ranges, no "up to" weasel words. The aesthetic moved from Loomgraph's purple-amber toward an editorial-tech feel: light mode, sans-serif primary with italic Garamond accents, a corner-bracket logo mark that doubles as a 3D extruded ornament in the hero with mouse parallax. The dashboard ships with backward-compatible auth - accepts both the new X-Citetome-Key header and the legacy X-Loomgraph-Key so nothing breaks during the cutover. 1,607 lines added across pricing, audit, and dashboard subpages in the same session. Built fast because the API was already real; the marketing site is just a window onto data that already existed.
EngineeringSystems
Agent-native commerce: why MealShift storefronts are built for AI shoppers, not just people
Deliveroo, Uber Eats, and DoorDash all login-gate their catalogues. An AI shopping agent hits a wall before it can even see what is for sale. MealShift just shipped the opposite: a sitemap, an AI-welcoming robots.txt, a machine-readable feed.json per storefront, and Schema.org JSON-LD keyed to the right vertical (Restaurant, GroceryStore, Florist, Pharmacy, Store) on every store page. An agent can discover the catalogue, read prices and stock, and transact through the ACP-style order API descriptor without ever touching a login form. This shipped in the same batch as the P0/P1 security fixes - Stripe webhook cross-tenant leak closed, refund stock restoration, safeguarding endpoint hardened, login brute-force throttled, ledger cross-tenant defence. Also landed: platform-wide activity tracking via a global NestJS interceptor, with a token-gated admin page doing 10-second live polling across four live tenants. The thesis: commerce is about to split into human-facing and agent-facing surfaces. Everyone is still building the first. A few of us are quietly building the second.
EngineeringSystems
Locking the name before writing the code: why Loomgraph got trademark-checked before the first commit
Named a new project today. Loomgraph: an agentic platform that auto-onboards sites, audits SEO, writes briefs, and weaves an internal-link graph across a portfolio. Sits on the ElektraOS runtime, pushes through the MealShift Commerce OS plugins. Before writing a line, I ran the three candidate names (Elektra Mesh, Synapse, Loomgraph) through PyPI, npm, GitHub orgs, and trademark search. Synapse was dead on arrival - Vertex Project Synapse on PyPI and Matrix Synapse homeserver own the name. Loomgraph: clean everywhere. Reasoning default is now three-to-five-year horizon: brand names, schema choices, and architecture are expensive to reverse. Cheap to verify on day zero, expensive to migrate on day one thousand. Research next: two parallel subagents mapping competitors (Surfer, MarketMuse, Clearscope, Frase, NeuronWriter, SEOwind, AirOps, Link Whisper, InLinks) and the open-source + free-API layer. The defensible wedge looks clear: portfolio-as-graph, not single-site optimisation.
EngineeringSystems
Migrating client portal encryption and fixing a two-character database crash
PageCrypt's CLI had a flag issue: the -o parameter silently wrote to a different directory than specified. Discovered it during QA. Migrated all 4 encrypted portals to StatiCrypt 3.5.4 - proper "Remember Me" support and shareable auto-decrypt links. Separately, the agent scheduler was crashing on line 619 because the content_calendar table was missing seo_keywords and meta_description columns. Two ALTER TABLE statements fixed it. Server restart, 8 priority agents executed within minutes. Also shipped: a Copenhagen-Accra network case page with fair pricing tiers calibrated to Ghana's creative economy ($25-200/month via Mobile Money), an interactive SVG network map spanning 6 cities, and a custom AI chat widget.
NeurodivergentEngineering
HTTPS enforcement, jsDelivr CDN, and why shipping is only 60% of the work
AYLA Maison (aylamaison.com) was showing "not secure" in browsers. Root cause: HTTPS enforcement was unchecked in GitHub Pages settings. The TLS cert was provisioned but not enforced. One checkbox fix. Then the management portal's product images were broken - cross-origin restrictions between two GitHub Pages domains. Routed all 9 product images through jsDelivr CDN (cdn.jsdelivr.net/gh/...) for zero CORS, zero cost, global caching. Also patched XSS vulnerabilities in the portal's innerHTML product rendering - replaced with safe DOM creation using textContent. Shipping is 60% of the work. The other 40% is verifying it works for the person using it, not just in your dev environment.
EngineeringSystems
I launched a jewelry brand, a puzzle game, and an API sandbox in the same session
aylamaison.com went live for a luxury jewelry brand in Casablanca. Custom domain, 4 DNS records, HTTPS, 9 products across 3 collections, and a dark luxury UI with particle effects. In the same session, The 40th Brick deployed on Render with Steam Deck gamepad support - 4 chapters, lobby, world map, achievements, all wired to controller input. And a client in London approved Phase 1 of his API sandbox after I fixed a duplicate schema key that was breaking Redoc. Three projects. Three industries. One system. That's what happens when your operating system handles the context switching for you.
NeurodivergentEngineering
Building a birthday gift that runs on a Steam Deck
Someone I care about turns 40 in three weeks. I could have bought flowers. Instead, I built a co-op puzzle game with 40 rooms, each one a memory or a challenge only she would understand. It runs in a browser, on a phone, and on her Steam Deck with full joystick and trigger support. There is an AI companion that gives cryptic clues but never the answer. The teaser page has a countdown, a riddle, and a whisper at the bottom that only appears when you hover. The neurodivergent brain doesn't send cards. It builds worlds.
SystemsEngineering
A portal where your LinkedIn becomes your entry ticket
I built a private network portal where I invite collaborators based on what they are actually good at. Not what they say on their CV - what an AI extracts from their LinkedIn profile and maps against open projects. JWT authentication, Kanban boards, skill matching scores, and a creative angle detector that identifies how someone thinks, not just what they know. The system matches people to tasks the way I would - by gut feeling backed by data. Every match comes with a score and a reason. Every action is logged. Nobody gets in without a password only I know.
EngineeringSystems
I encrypted a client's entire project portal in 3 seconds
A founder sent me a 10-page API PDF on WhatsApp. Two hours later, he had a password-protected developer portal with all 11 endpoints documented, an OpenAPI 3.0.3 spec, and an AI assistant that reads the page state and answers his questions about build progress. The encryption happened in 3 seconds. The architecture decision log, the dual-clock time tracking, and the context-aware chat took longer. But the point is: the hard part isn't building the docs. It's building the system that makes the docs look like they took a team of 5 and a month to produce.
Engineering
64 pages in 4 languages and I didn't translate a single word manually
I launched elektraos.dev in English, then deployed Danish, French, and Italian versions within the same session. 48 additional pages. Each one with correct hreflang tags, localized nav, market-specific pricing in local currencies, and browser detection that suggests switching. The translations preserve tone: Danish gets the informal "du" form. French uses "vous." Italian uses "Lei." Each market has its own intelligence agent monitoring local regulations, business culture, and industry news. Not because I speak four languages. Because I built a system that does.
SystemsNeurodivergent
46 agents, 4 markets, and a security audit that found 4 critical issues
Today I registered 4 market intelligence agents (Denmark, France, Italy, EU regulatory watchdog) and immediately ran a security audit on the entire system. Found 4 high-severity issues: exposed API docs, overly broad CORS, and leaked architecture details through the health endpoint. Fixed all four in the same session. The audit agent that found the problems and the market agents that monitor compliance now run alongside the 41 agents that were already active. Every new capability adds a new attack surface. The neurodivergent brain spots the patterns. The system catches the rest.
NeurodivergentSystems
Nothing is ever truly shipped
The neurodivergent brain doesn't do "done." Every project is a living system. Every feature triggers three new ideas. Every deploy opens a new iteration cycle. Most people see this as a problem. I built an operating system around it. ElektraOS tracks 6 active projects simultaneously, each one always improving, always evolving. The trick isn't forcing yourself to finish. It's building infrastructure that turns perpetual iteration into compounding output.
Engineering
Why I automated my own documentation
I set up a pipeline that logs every engineering session automatically. An agent tracks what was built, measures working time, and publishes structured entries to the build log every 2 hours. No manual writing. No editing. The numbers are real, the timestamps are real. If I have to choose between documenting and building, I'll always choose building. So I made the documentation part of the build process itself.
EngineeringNeurodivergent
I used to work at companies that raised millions. Now I build their solutions in days.
WakeUpData built a data feed management platform. Contractbook built contract automation. Both raised millions. Both took years. I sat inside both companies. Today I can prototype equivalents in a weekend. Not because the problems got simpler. Because the tools got exponentially better. The gap between "idea" and "production" collapsed. The question isn't whether you can build it anymore. It's whether you can think clearly about what to build.
Systems
Building a thumbnail pipeline that works across 3 brands
The hardest part of AI image generation isn't the generation. It's consistency. Same face across every thumbnail. Same brand feel across different platforms. Same quality whether it's YouTube 16:9 or TikTok 9:16. I built a system that solves this with face locking, JSON-structured prompts, and a multi-brand asset library. Each approved output becomes a future reference. Consistency compounds.
EngineeringSystems
184,000 lines of code and counting
ElektraOS hit 184K lines. 41 AI agents. 624 API endpoints. 59 database tables. These aren't aspirational numbers. These are verified from the codebase. The system runs 16 cron jobs autonomously, orchestrates 5 agent teams, and handles everything from content generation to lead scoring to invoice parsing. It started as a personal tool. It became the proof that one person with the right infrastructure can operate at the scale of a small company.
This blog is automated. Posts are generated from live engineering sessions and published through a multi-agent pipeline. 5 agents coordinate across vault, API, and deployment layers.
New posts publish alongside the build log. Building is writing.