WikiDesignCo · Looikos

Project: WikiDesignCo (WDC)
Looikos cluster: Infrastructure & Agent Platforms (the substrate / data-platform lab)
One-line: Andy's Skunk Works for data platforms. The internal laboratory of "crazy wonky shit" where the metagraph world-model gets built, then rented to the market as Forward Deployed AI Engineering on a flat retainer so an operator never has to make the $400K dual hire.
Status: In-build. Wave A (scaffold + Library, 14 articles) live. Wave B platform build STARTED 2026-07-03: team wdc-platform-v1 in flight (separate GCP project, workspace-scoped Convex layer, walking-skeleton gate before workspace fan-out). The ContentFactory live instance is the live customer-facing receipt: the dealership client pays $2,800/month right now. See §7b.

1. What it is (the one-paragraph truth)

Picture the operator who knows AI is supposed to change their business and can't make it stick. They paid a developer to build an internal knowledge tool and got a six-week flaming pile of brittle code that nobody will touch now. They looked at hiring the person who could do it right, a Forward Deployed Engineer (FDE, an engineer who works embedded inside the customer's business), and found that senior FDEs in major US tech hubs command total comp well into the mid-six figures, often exceeding $300,000 and reaching toward $400,000-plus at the top end (per self-reported comp data, so directional rather than an official band), and that the role is widely reported as one of the hardest customer-facing engineering profiles to hire, gated behind a budget they don't have. So they keep loading a 200,000-word corpus into every agent call, burning money on redundant processing, watching the agents improvise and hallucinate generic garbage, and telling themselves they'll fix the knowledge problem later. Later doesn't come. That operator is who WikiDesignCo is for, and naming their exact bind is the whole pitch.

WikiDesignCo is the internal data-platform laboratory at the center of the Looikos brands, and it rents the same capability to that operator as Forward Deployed AI Engineering on a flat monthly retainer. In his walkthrough Andy compares it to the research department an institutional quant firm keeps, "this fucking laboratory of financial engineering" that runs the crazy wonky experimental work nobody else will touch. WikiDesignCo is that lab for data platforms, and it's his Skunk Works: people know it exists, it stays mostly internal for now, and enterprise offers and accessible products come over time. What the operator rents is a knowledge layer built on a metagraph (a knowledge graph that also records where each fact came from and when it was true), with precise retrieval served to AI agents through MCP, the standard protocol agents use to reach tools and data. It comes wrapped in an operating model where one operator-architect plus an agentic stack delivers what a fourteen-specialist publishing house plus a dual-hire FDE team would deliver in-house. The customer never sees the GraphRAG, the temporal validity windows, or the contradiction engine. They see a flat number per tier and a deliverable that holds up. The live receipt is a client paying the $2,800/month part-time tier right now, and the proof-of-platform artifact is the Library, where each 10,000-plus-word piece is itself a sample of the work WikiDesignCo sells. Each article is the demonstration, so the brand makes its case by doing the work in public.

Andy's words (verbatim): "We've got Wiki Design Pro, which is like the central data platform. This is an internal asset. I don't actually intend to release this in a major public format. It was kind of like the equivalent of an institutional quant firm where they might have a research department that comes up with all sorts of crazy wonky shit. They basically really wonk this fucking laboratory of financial engineering. We have the same thing, in this case here that's Wiki design code. Wiki design code is that but for data platforms. What [the agent] harness is for all of this modern AI architecture, AI infrastructure, whatever you want to call it. Wiki Design company is the laboratory where it's my version of Skunk Works. People know it exists. Eventually I'll have enterprise offers for it and probably even some accessible stuff over time. But point is that this is where things are experimental right now."

Reading between the lines. Three things sit compressed in that quote, and WikiDesignCo's business brief confirms each one past the point of inference.

The first is what the Skunk Works framing buys, which is a posture, not a hedge about maturity. A Skunk Works is the place you run the experiments the main line can't afford to fail in public, staffed by people trusted to work without the usual approvals. Lockheed's built the U-2 and the SR-71 in a fraction of the normal program time because it was insulated from the bureaucracy. WikiDesignCo is that for data platforms: the lab where the metagraph world-model, the temporal-validity database, the contradiction engine, and the agentic content compiler get built and stress-tested before any of the other thirty-plus brands depend on them. When Andy says "this is where things are experimental right now," he's describing the operating model. The lab ships the dangerous, novel work first so the rest of the ecosystem can rent it as settled infrastructure, and the customer who rents it never carries the experiment risk.

The second is that WikiDesignCo is ContentFactory V2. ContentFactory was the first live stress-test of the whole idea, and its live instance is the customer-facing receipt from the opening, the paying client that proves the model converts. WikiDesignCo is the production-grade rebuild underneath it, with everything ContentFactory learned it needed built in from the first commit rather than retrofitted: a metagraph, a hexagonal ports-and-adapters architecture, dependency layering, durable execution through Inngest, observability through LogFire, and property-test evaluation through Hypothesis, all from day one. ContentFactory then becomes a customer-facing GUI that runs on top of WikiDesignCo infrastructure, and the downstream Constellation brands (FreelanceBuddy, Social Storyboard, Constellation Media, Node Foreman, Quant Scientist, Depths of the Void) inherit the trust WikiDesignCo establishes. That inheritance is the strategic point. Among forty brands, WikiDesignCo is the floor the others stand on, which is why it's researched first and ranked as foundational substrate rather than a leaf.

The third is the shape of the moat, which is three walls stacked in sequence. The first wall is metagraph world-modeling, which treats provenance (where each fact came from), temporal validity windows (when it was true), and a contradiction engine as first-class structure rather than bolted-on metadata, so the knowledge layer knows what's true, when it was true, and where it disagrees with itself. The second is operator-plus-FDE compression: one operator-architect plus the agentic stack does the work of a fourteen-specialist publishing house plus a dual-hire FDE team, which is where the accessible pricing comes from without the quality dropping. The third is credit abstraction: the customer-facing interface is priced in credits with an AAA-game look, a Diablo-style budget-tier selector where the customer self-selects a spend ceiling, which makes profitability invisible to the customer while making the customer's own budget predictable. Test which wall holds against the seven sins, a checklist of seven ways an analysis fools itself (look-ahead, survivorship, overfitting and the rest). The seductive answer, flat-retainer pricing that absorbs API variance, is a look-ahead trap: it's a real wedge today and copyable the moment tokens commoditize, so scoring it as the moat scores a fiction. The durable moat is walls one and two, the verticalized knowledge fabric that turns onboarding into mapping a customer's data onto an existing ontology rather than rebuilding from scratch, plus the compression that lets one person ship enterprise-grade output. Andy's framing and the cold outside read land on the same two, and that agreement is the strongest signal the moat is sited correctly rather than wished into place.

WikiDesignCo is the live instance of the world-model that Harness V2, the next version of Andy's agent harness, is built to serve, and the software every other brand resells is built on the feature factories this lab proves out first.

3. The three-angle valuation (the core of a self-standing brand)

3a. Finance (credit and capital access)

Andy reads every brand the way a market maker reads a target: fundamentals plus technicals plus live sentiment, verified against the bank account, not the dashboard. For WikiDesignCo the questions are what economic activity it throws off, how that activity converts to credit and capital access, and what the brand is worth to an acquirer.

The economic activity is recurring retainer revenue against a known cost of delivery. The retainer tiers are real and live: $2,800/month part-time, $4,000/month full-time, $8,000/month real-time, monthly exit, quarterly lock. The dealership client sits at the part-time tier today, and the internal cost of delivery at peak runs roughly $5,300/month in credits that WikiDesignCo absorbs. That looks upside-down for a single early account, and it's meant to. The markup math runs aggregate_cost x 1.25 reserve x 1.65 grandfathered = 2.0625x for early customers and x 1.8 base = 2.25x for standard customers, trending toward 3x to 5x cost-plus-reserves over the three-to-six-year horizon as the platform matures and upstream token prices fall. The arbitrage is explicit: the gap between today's cost of delivery and tomorrow's commoditized floor is a finite window, and every customer onboarded at the current markup funds the subsidized-credit moat the next customer rents. The independent market read backs the mechanism. GraphRAG and knowledge-based retrieval run roughly ten times fewer tokens than naive vector-only RAG on complex queries and cut cost by around 67% at scale, which is what makes a flat retainer that absorbs usage variance margin-safe rather than reckless.

The credit story for an infrastructure brand runs on the quality of that recurring revenue. WikiDesignCo isn't a heavy advertiser, so the pull that ad spend gives some of the agency brands with their banks is muted here. What WikiDesignCo offers a lender instead is the cleanest collateral a young software business can have: contracted monthly recurring revenue with quarterly locks, a measurable and improving gross margin, and net revenue retention that should sit in the best-in-class band as customers expand from part-time to full-time to real-time tiers and add platform usage. Best-in-class vertical-AI and SaaS-plus-services businesses run blended gross margin of 60% to 70% and net revenue retention of 115% to 125%. Revenue-based-financing desks and venture-debt lenders lend against recurring revenue of that quality, and the predictability of the flat-tier model plus the audit-stage quantification means the forward revenue is forecastable enough to factor. The capital path is the standard one for a brand of this profile: private venture and venture-debt early, with the public path open only if the brand consolidates several Constellation surfaces into one reportable platform entity.

Looikos sets $10M as the minimum value each of a brand's three angles should reach, and the M&A and valuation read shows why that number is a floor here, not a ceiling. The category comps are strong and named. On the infrastructure side, Pinecone raised a $100M Series C in 2023 at a reported ~$750M post and a $200M Series D in 2024 widely reported above $2B; Glean raised a $100M Series B in 2022 at roughly $1B and a $200M Series C in early 2024 at roughly $2.2B; Neo4j, the cleanest graph-backed-data-platform comp, raised a $325M Series F in 2021 at more than $2B, the largest database round in history at the time, at an implied ~13x to 20x ARR. Those are the ceiling comps, in the same category as WikiDesignCo's metagraph-plus-RAG layer. What pulls a real number down is the services discount: pure product and infrastructure trade at roughly 8x to 14x revenue when growth and retention are strong, while AI-implementation services with limited IP trade at 1x to 3x revenue, so a business that is 60% to 70% recurring platform revenue and 30% to 40% services lands in a defensible 5x to 9x blended band. Run that against even a modest steady state. One hundred to two hundred and fifty retainer customers across the tiers, the standard service floor Looikos plans every brand around, produce annual recurring revenue (ARR) in the low-to-mid eight figures; at a 5x to 9x blended multiple that is a $50M-to-$150M enterprise value for the service-and-platform angle alone, which is why Andy holds $10M as the floor for one angle rather than the target for the whole brand.

The market maker's three-level read ties it together. The fundamentals are the unit economics already covered: contracted ARR, a 60%-to-70% blended margin protected by the GraphRAG token savings, and an expanding net revenue retention. The technicals are the funnel and retention mechanics, where flat tiers with a monthly exit and a quarterly lock are engineered for low churn and tier expansion. The live sentiment is the strongest tailwind of all: Forward Deployed Engineer postings rose more than 800% between January and September 2025, Salesforce alone is building a team of a thousand FDEs, and the role is one of the scarcest and most expensive in enterprise tech. A brand that sells that scarce capability at an accessible price is reading a market where demand is revealed, supply is constrained, and the narrative is moving its way.

3b. Software (the interface stack)

The software angle is the one every other Looikos brand resells, so it carries the most strategic weight. WikiDesignCo's product is a metagraph-backed knowledge platform exposed through one core and many thin interfaces, built on the hexagonal ports-and-adapters pattern so the same operations reach every interface without being duplicated.

The core is the metagraph backend. Convex holds the source of truth, Neo4j plus Graphiti hold the temporal concept-graph with its validity windows and contradiction structure, Qdrant holds the vectors, and Typesense holds the full-text index. None of those stores re-implements the graph operations; each is an adapter over a core that never imports a transport. That one rule lets a single knowledge layer present itself at once as an HTTP API, an MCP server, a CLI, a web UI, and an agent interface, with one source of truth behind all of them. The market read shows what that's worth. Glean, Vectara, Sana, and Credal each sell a slice of this and refuse the rest, and the pure knowledge-graph vendors such as Stardog and Oxford Semantic are too heavyweight to onboard a small business. WikiDesignCo's software angle is the assembled whole that none of them sells as one thing.

Each interface maps to a revenue line, which is the point of building them as separate adapters. The MCP server sells agent access: downstream Constellation brands and external agents query exactly the knowledge they need through MCP rather than loading a 200,000-word corpus, a roughly 90% reduction in context usage, and that agent access is metered in credits. The CLI and the API support a credit-based and subscription program for programmatic consumers. The web UI, which is the Library today and becomes the Platform when Wave B (the production platform build) lands, is the SaaS subscription for human operators. The customer-facing pricing layer is a piece of software with a deliberate look: it's priced in credits and styled on AAA-game monetization, with a Diablo-style budget-tier selector where the customer self-selects a spend ceiling between hundreds and tens of thousands of credits. That credit abstraction does real economic work. It hides the platform's profitability from the customer while making the customer's budget predictable, which resolves the hardest problem the market read surfaced: flat all-you-can-eat retainers are brittle if one client hammers an internal chatbot and collapses the margin, so the credit ceiling is the mechanism that keeps the flat-feeling price margin-safe.

Underneath the surfaces, the platform decomposes into feature factories with clean domain boundaries, each domain maintained largely automatically by its dedicated agent harness. WikiDesignCo's factories are legible from the repo: ingestion (web crawl, document processing, chunking, embedding, indexing, inherited from the Archon-forked RAG core), the Content Compiler agentic stack that turns ingested knowledge into in-brand assets, retrieval and the AndyDataBot surface that serves precise context, and the metagraph world-model layer that holds provenance, temporal validity, and contradiction as first-class structure. Each factory is a set of agent harnesses plus a gateway harness specialized for its domain, so the already-domain-specialized agents submit the work and quality rises rather than falling as scope grows. That's why the Harness V2 build matters to the finance and service angles too: the software is built from custom, modular harnesses that combine into feature factories, so each one is built once, maintained cheaply, and sold three ways.

The bundled-tools subscription is the part the customer experiences as simplicity and the operator experiences as managed complexity. The software line item on a WikiDesignCo invoice bundles roughly twenty-four tools (LogFire, LangGraph, PydanticAI, Graphiti, Neo4j, Convex, Vercel, Higgsfield, Suno, ElevenLabs, Veo, Sora, Inngest, Clerk, Qdrant, Typesense, Remotion, Three.js, Framer Motion, GSAP, Payload, the Claude Agent SDK, the Vercel AI SDK, and the Hypothesis property-test stack) with the management fee baked into the markup. The bundle works like a construction project manager running twenty-four subcontractors: the customer hires one number and one accountable party, and the platform absorbs the coordination and the cost variance behind it. The product's value is that the customer never has to assemble or reconcile the stack, which is the same thing the FDE hire offers at fifty times the price.

3c. Service (premium-at-accessible boutique delivery)

The service angle is where WikiDesignCo touches a paying customer today, and it's the angle the whole brand was reverse-engineered from. The model is Forward Deployed AI Engineering rented as a boutique retainer, premium quality at accessible pricing, sold as one flat number per tier.

The retainer structure is live and specific. There are three tiers, $2,800/month part-time, $4,000/month full-time, $8,000/month real-time, with monthly exit and an option to lock the monthly rate on a quarterly contract. The customer sees two line items and nothing else: the operator retainer, which is labor plus accountability, and the software subscription, which is the bundled platform with the management fee in the markup. The alignment mechanism that makes the flat price hold across an engagement is audit-stage quantification. At engagement start, an audit locks the ingestion intensity (light, standard, or heavy) and the deliverable shape (per document, per corpus, per sprint), and the platform quantifies the deliverable price against that audit. Because the metagraph stack is extendable, modular, and maintainable, the audit estimates hold, so the customer experiences zero mid-engagement pricing surprise. The audit is the repo's design answer to the failure mode the market read flagged: retainers that promise unlimited usage collapse unless the scope is engineered tightly, and the audit-plus-credit-ceiling pairing is how WikiDesignCo presents a flat predictable price while staying margin-safe behind it.

The target operator is the standard Looikos buyer, the owner of a business under twenty-five people who has mastered a craft (real, durable expertise nobody else can copy) but can't scale it. For WikiDesignCo specifically that resolves to the knowledge-heavy and content-heavy SMB or agency: the operator whose real value is locked in documents, email, a CRM, and their own head, who needs FDE-grade knowledge engineering but can't hire a $350,000-to-$500,000 Forward Deployed Engineer and wouldn't get one to take a $2,800 retainer if they could. The market read confirms the gap is real and unserved from both sides: product-company FDEs are too expensive, are tied to one platform, and won't touch accounts under $10K, while the cheap AI-automation agencies ship shallow Zapier workflows and never build a real knowledge platform, knowledge graph, or governed ingestion. WikiDesignCo sits in that gap, with the infrastructure depth of the enterprise vendors and the price accessibility of the agencies.

The compression is what makes premium quality at an accessible price add up. At the part-time tier the customer pays $2,800/month and buys the equivalent of a fourteen-specialist enterprise-grade publishing house, roughly $2.5M to $3.5M a year fully loaded, plus a dual-hire FDE team at roughly $2.4M a year fully loaded, for a total in-house equivalent near $5M to $6M a year, or $416,000 to $500,000 a month. Against the $2,800 the customer pays, that's a compression of 149 to 178 times. The market read independently corroborates the mechanism behind the number: the durable edge is decomposing the FDE role into a small core of senior lead architects who design the playbooks plus a templatized infrastructure that delivers FDE-grade outcomes without staffing an army of $400,000 engineers. WikiDesignCo's version of that split is one operator-architect plus the agentic stack, the compression moat from the seed reading seen from the service side.

The angle bottoms out around $1M/month at the standard Looikos count of 100 to 250 retainer customers. It scales well above that as customers move from part-time to full-time to real-time and add platform usage, which is the expansion in net revenue retention that the finance angle counts on. The commodity work beneath the premium engagements (routine content production, basic automation) gets partnered to the sister affiliate network of specialists, so service at scale is itself a network rather than a headcount problem, and the customer relationship runs on the shared floor, the Looikos customer-success model in which senior people rotate through one room with AI agents listening in. The relationship with the customer is the irreducibly human part, and it's the part the whole agent-native stack exists to make one person capable of delivering at portfolio scale.

4. The personas (5+, modeled to world-experience depth)

Six personas speak here in the first person, each in language close to how these people talk. Each one drills past the surface complaint ("I need a knowledge tool") to the buried shame several layers down ("I'm winging it and praying," "I built a graveyard of scripts out of ego," "I'm silently failing at operations"), because the operator who names that deepest pain owns the solution in the buyer's mind before any feature gets mentioned. They're the human side of what Andy sells as Intelligence Infrastructure, his service line for market research, data analytics, business automation, and CRM. The same trauma he names in his service writing (research done to tick a box, analysis paralysis dressed up as rigor, the data graveyard that cost six figures) is the pain WikiDesignCo's metagraph absorbs. The pain language in these personas is provisional rather than quoted: the voice-of-customer research returned constructed but realistic phrasings instead of verbatim citations (the source said it couldn't surface exact thread quotes), so each phrase is representative voice, not a documented quote. The patterns are corroborated where the business brief confirms them (the six-week flaming pile is in the brief verbatim) and where the market read confirms the structural situation (the FDE gap, and the majority of small businesses still running on manual processes). The personas lean toward the negative emotions, because that's where these people live.

P1. The master-craftsman who cannot scale (the canonical target)

I'm amazing at the work and absolute garbage at everything around it. I built this thing with my bare hands and I'm proud of it, and I spend my day doing invoices, chasing people, and fixing dumb little fires, so by the time I get to the work I'm supposed to be known for I'm fried. It feels like I'm babysitting my business instead of running it. I see these AI-first shops pumping out work like a factory and I'm over here duct-taping spreadsheets together just to keep up. Every week there's a new AI tool and some guy screaming "adapt or die," and I'm still trying to write my own SOPs.

Last month a client asked how I'm using AI in my process, and I said we're exploring options, which is a polite way of saying I'm winging it and praying. That's the humiliating part. I know I'm good, but it's starting to feel like being good isn't enough anymore. I got here by optimizing for the craft for twenty years and never building the systems, because the craft was the thing and the operations were always going to get handled later. Later never came. Getting out means systematizing knowledge I've never written down, and most people like me fail at it because documenting it is so overwhelming that we abandon it halfway and go back to doing it manually, which is the trap. Staying stuck means living with the slow certainty that one bad month and my clients realize they can get eighty percent of what I do from some platform for a fraction of the price. Getting out costs me an admission of how messy it's been the whole time. WikiDesignCo is for the operator who's ready to stop being the bottleneck without becoming irrelevant, by renting the FDE-grade systems that turn the craft into something that scales beyond their own hands.

P2. The technical founder who built their own RAG and abandoned it

I did the classic engineer thing. We burned six weeks rolling our own lightweight RAG system because how hard can it be, and we shipped nothing, and now I have a graveyard of half-baked scripts I refuse to open. I spent a month tuning embeddings and chunk sizes for a tool nobody outside the dev team even asked for. The whole thing is so brittle that any schema change means re-indexing everything and praying it doesn't silently corrupt. Our knowledge base is now half in Notion, half in a janky Postgres table with embeddings, and half in people's heads, which is three halves.

It was pure ego. I wanted to say "we built our own RAG stack" instead of "we pay a boring monthly fee for something that works." Leadership now sees me as the person who chases shiny objects, and I've already spent my credibility for the next AI initiative I propose. What got me here was the senior-engineer urge to build the database myself, the conviction that a managed solution felt bloated and we could do it leaner, and a refusal to pay for the thing that solved eighty percent of the use case out of the box. Getting out means swallowing the identity hit of becoming someone who rents the plumbing instead of building it, and that's harder than the engineering. Most people fail right there, because the blocker is the build-everything-in-house identity rather than the technology, and most engineers will defend that identity past the point where it's hurting the company. Staying stuck costs the next six weeks, and the customer-facing feature that never shipped because three devs were bikeshedding indexes. WikiDesignCo is the off-ramp that lets the founder keep the engineering pride for the product and rent the knowledge infrastructure that was never the differentiator.

P3. The SMB owner drowning in fragmented knowledge and content operations

Everything is everywhere and nowhere at the same time. We have docs in Drive, SOPs in Notion, the real process in the brain of a manager who's always in meetings, and every time someone asks "where's X?" I die a little inside because the answer is basically "good luck." We rewrite the same thing over and over because nobody can find the previous version or trusts that it's current. Our content strategy is panic: produce something, throw it in a random folder, repeat. We're on a treadmill where we're constantly producing and nothing gets reused, and it's exhausting.

I open our shared drive and feel instant decision fatigue. Is it in Marketing, Old Marketing, Archive, To Sort, or Misc? If I got hit by a bus, half the company's knowledge would disappear with me, and I'm ashamed of how much time we waste asking the same questions in Slack because nobody can find the answer. Every fix I tried turned into another abandoned folder or tool, so I stopped trying to fix the system and absorbed the chaos into my memory, which made me the single point of failure. The way out is a knowledge layer that holds the institutional memory and serves it back precisely, which is the problem WikiDesignCo's core was built to solve: precise retrieval instead of dumping everything into every request. Most people never get there because the chaos is invisible from outside, so there's never a forcing crisis, just a slow exhaustion that never quite tips into action. Staying stuck means burnout and the quiet fear that real businesses have their act together while I'm silently failing at operations. Getting out means letting a system hold what I've been holding in my head.

P4. The senior practitioner whose expertise is trapped in their head

I keep getting asked to just document my process, and I don't know how to explain that my process is twenty years of pattern recognition, not a checklist. So much of what I do is gut feel backed by scars from all the times things went wrong, and you don't put that in a wiki page. Every time I try to write it down I hit a thousand edge cases and exceptions and I give up halfway through. I can tell in five minutes that something is off, but if you ask me why I would need an hour to reconstruct the reasoning. People say "we just need to clone you" like it's a joke, but what I hear is "you're the bottleneck."

I'm tired of being the person everyone pings to unblock things, and I also don't trust that the work gets done right if I step back, so I'm trapped between resentment and control. The expertise built up invisibly, one corrected mistake at a time, and by the time anyone needed it transferred it had already become tacit and out of reach, the kind of knowledge the person carrying it can't fully see. I'd need a system that extracts the reasoning by interrogating the work instead of asking me to introspect cold, and that's what WikiDesignCo's metagraph and agentic ingestion are built for: capturing the entities, the relationships, and the provenance of expert judgment rather than demanding a hand-written manual. Most people like me never do it, because documenting the expertise feels like manufacturing my own replaceability, so the move that feels safer is always to keep it in my head, and that fear quietly wins. Staying stuck leaves me a fragile single point of failure the whole team is exposed to. Getting out takes the courage to believe that externalizing the judgment makes me more valuable as the person who designs the system, not less.

P5. The in-house AI champion under pressure to show ROI

I feel like I got sold a sci-fi demo and handed a glorified autocomplete with a dashboard. Leadership saw a slick vendor demo and now I'm on the hook to transform the business with what's essentially an API wrapper with a logo. We ran a pilot, everyone played with the chatbot for a week, and then it died quietly, and now I'm the person who wasted that budget. The tool works in the sense that it doesn't crash, but nobody trusts it for real work. Half the tools we evaluated turned out to be thin wrappers over a foundation model with some logging, and I wish I had pushed harder on due diligence.

I'm stuck between skeptical engineers who think it's snake oil and execs who want a case-study-worthy success next quarter, and I'm embarrassed because I was the one pushing hard for this, so it feels like I fell for the hype deck. The pressure to be the innovation champion ran ahead of the diligence, and the vendor's promise that it would just plug in turned into three months of SSO purgatory, permissions hell, and weird edge cases with our data. What I need is a partner who owns the last mile (the workflow redesign, the data plumbing, the change management) instead of a tool that hands me a component and walks away, and that's the line the market read draws between FDE-grade work and a shallow wrapper, the gap WikiDesignCo occupies. Most people in my seat fail because a disappointing pilot poisons the organization against AI entirely, so the next attempt fights both the original problem and the scar tissue. Staying stuck makes me the face of an expensive, visible failure, with my career tied to results I don't fully control. Getting out means admitting the first vendor was a wrapper and choosing depth over another demo.

P6. The enterprise evaluation lead deciding build-FDE-in-house versus rent

I'm the one who has to recommend whether we hire the Forward Deployed Engineer or rent the capability. The hire is real and the number is brutal: total comp clears three hundred and fifty to five hundred thousand, the role is one of the hardest to fill in enterprise tech, and postings for it went up more than eight hundred percent in nine months, so even with the budget approved I'm competing for a person who has so many high-paying offers coming at them that I probably lose the bidding war. If I do land them, they're one person, they take months to ramp, and if they leave the capability leaves with them.

I own this decision. If I build in-house and the hire walks in a year, I own that outcome, and if I rent and the vendor underdelivers, I own that one too, so the decision feels like choosing which way to be exposed. Leadership decided AI implementation is the bottleneck, which it is, and handed me the build-versus-buy call with no obviously safe answer. Deciding well needs a rented option that's credibly FDE-grade rather than agency-shallow, with the cost variance absorbed so the budget is predictable, and enough depth that it doesn't become another disconnected pilot. Most evaluators stall because the enterprise-grade vendors won't take a mid-market account and the cheap agencies can't do the real work, which leaves them stuck between two non-fits in the unserved middle WikiDesignCo targets. Staying stuck keeps the bottleneck a bottleneck while the quarter burns. Choosing means trusting a boutique to do what a $400,000 hire would, which the compression math and the live dealership receipt are built to make credible.

5. The world model (run the PST framework)

The six personas share one underlying story, and modeling it as a single loop of suffering is what turns them from demographics into PST, Andy's framework for reading a buyer's Problem, Story, and Transformation. It runs in four moves: echolocate the world, locate the Problem, reconstruct the Story, and design the Transformation.

Echolocate the world. The buyer lives inside a market that's itself in motion, and it's read here the way an institutional M&A firm reads a target. On one side are the buyer's customers, the people who pay them, whose expectations are being reset weekly by what AI now makes possible, so the buyer feels the floor rising under them through their clients. On the other side is the supply of capability, and the FDE who could fix this is one of the scarcest and most expensive people in the labor market, for the reasons the finance read laid out, so the obvious solution is structurally out of reach. Between those two pressures sits a loud, untrustworthy tooling market: enterprise platforms that won't serve them, agencies that ship shallow wrappers, and a constant stream of demos that overpromise. Mapped as a graph, this world is dense with relationships that don't fit in rows: the buyer's knowledge, their clients' shifting expectations, the talent market, the vendor noise, the money and the blame flowing between them. WikiDesignCo is built on a metagraph because that's the shape of the customer's world, and you can't model it as a spreadsheet.

Locate the Problem (the cycle of suffering). The pain that arrives is concrete and recurring: knowledge is fragmented and unscalable, the craft can't leave the operator's hands, and the one time they tried to fix it themselves it became a brittle mess. In response a fear gets installed, and most of these buyers carry a bad mix of them: fear of irrelevance (being lapped by AI-first competitors), fear of humiliation (the six-week flaming pile, the client who asks how they use AI and hears "we're exploring options"), and fear of being conned again (the wrapper with a logo that died quietly and took the budget with it). Those fears drive avoidance, and avoidance produces the unfavorable outcome: the systems never get built, the operator stays the bottleneck, the pilot stays disconnected. The outcome produces shame, the belief that I'm bad at this rather than that I did a bad thing: secretly chaotic after all these years, the engineer whose ego cost the company two sprints. The shame is unbearable, so it gets buried under cope: blame the hype cycle, blame the vendors, blame the pace of change, blame the LinkedIn guy screaming adapt or die. The one forbidden move, the red line, is accountability, because accountability means turning around and admitting that the stalled platform and the trapped expertise came from fears they let drive the decisions, not from the market. The refusal opens a blind spot, the blind spot produces the next disadvantageous action (another abandoned folder, another in-house rebuild, another demo-driven pilot), and the loop closes and compounds. Almost every WikiDesignCo buyer is stuck at this point, and the content has to meet them there, in the denial and the shame, not in the clean future state.

Reconstruct the Story. Under the loop sits some version of the belief "we should be able to handle this ourselves." For the technical founder it's the build-everything-in-house identity, for the master craftsman it's "the craft is the thing and operations are beneath it," and for the senior practitioner it's "my judgment can't fit in a template." The chain that built it runs the way the framework describes. Being rewarded again and again for individual mastery hardened into a belief that mastery is the whole game, the belief drove actions (optimizing the craft, refusing the managed tool, hoarding the expertise), the actions produced results, the results became habits, and the habits anchored into an identity. At the origin there's usually a wound around worth. Somewhere the person learned that their value is the thing only they can do, so anything that externalizes or systematizes that thing reads as a threat to their worth rather than a multiplication of it. Most of them run from that wound. It's where documenting the expertise feels like manufacturing their own replaceability, and where renting the plumbing feels like admitting they were never as indispensable as the identity required. PST's archive of 130 emotions and the Hawkins scale (a ranking of emotional states from shame upward, with courage as the line between the destructive and the constructive) are used here descriptively, to locate where each persona sits: shame, fear, and pride live in the destructive band below the courage line, and the whole loop of suffering runs on them.

Design the Transformation (the cycle of growth). The bridge across has to be one the buyer can cross without feeling mugged, and it hinges on courage, the point that separates the destructive band from the constructive one. The first step is truth, which WikiDesignCo's content surfaces gently: the craft was always the differentiator, not the knowledge infrastructure, and renting the plumbing frees the craft rather than diminishing it. The second is responsibility, owning the reaction rather than the circumstance: the buyer didn't cause the talent shortage or the vendor noise, but they own whether they keep letting the fear of looking dispensable drive the decision. The third is healing, which hurts the way relearning to walk hurts, because externalizing twenty years of tacit judgment or admitting a failed build means tearing through the identity knot that the expertise is the self. The fourth is forgiveness: letting go of the old verdict, forgiving the six weeks and the ego and the could-have-bought-it, and having the humility to learn from it. That opens the buyer's eyes to a new truth, that the architect of a system plays a larger role than its single point of failure. WikiDesignCo's offer is calibrated to that bridge. The flat retainer removes the financial fear, the audit-stage quantification removes the surprise, the live dealership receipt and the compression math remove the will-this-be-another-wrapper fear, and the operator-plus-agentic-stack model lets the buyer cross from being the bottleneck to being the person who designed the thing that scaled. Most of the content lives in the negative band because most of the audience lives there, with the growth cycle shown as the reachable other side. That's the approach Andy's Echolocation and Mirror Ocean articles describe, applied to this customer: model the world fully, name where they're stuck, and offer the way across.

6. Competitive and market read (the alpha / third door)

The competitive field is crowded at the edges and empty in the middle WikiDesignCo occupies. It's mapped here by what each player does, what each refuses to do, and where the third door is, the option neither end of the market will take.

Who else does this, and what they won't do. There are three clusters of competitors plus one substitute. The enterprise knowledge and RAG platforms (Glean, Vectara, Sana, Credal) each sell one slice and refuse the rest. Glean does permission-aware enterprise search well but doesn't redesign workflows, doesn't act as an FDE, and doesn't own the customer's business ontology beyond what search needs. Vectara delivers quality managed retrieval but still requires the customer's own engineers to design the application, the prompts, and the business logic, so it's a component, not an outcome. Sana is knowledge management and learning, not end-to-end process automation, with limited professional services. Credal solves secure access and governance but is infrastructure rather than outcomes. None of them says "for a flat monthly number, we'll own your AI knowledge infrastructure and deliver specific outcomes," and none of them serves small or mid-market businesses at all; they earn from platform subscription plus usage, not from absorbing the variance. The knowledge-graph and GraphRAG vendors (Neo4j, Stardog, Oxford Semantic, Writer's graph-RAG, IBM GraphRAG) own the technology WikiDesignCo's metagraph is built on, but they sell tooling to enterprises that already think in ontologies and already have data teams; they won't sit with a fifty-person company and design its ontology, and they're too heavyweight and too expensive to onboard a small business. The AI-automation agencies sit at the opposite extreme: hundreds of small shops shipping $500-to-$2,000 retainers on Zapier and no-code stacks at roughly eighty-five percent margins, who understand SMB workflows but never touch hard data plumbing, never build a real knowledge platform or graph, and never do genuine FDE work (process redesign, change management, owning a KPI over a long horizon). The fourth competitor is the substitute itself: the in-house FDE hire, which is the thing the buyer would do if they could, and can't, because the role clears $350K-to-$500K, is one of the hardest to fill in enterprise tech, and the candidates have so many offers that even a funded employer loses the bidding war.

The third door. Alpha, the investor's word for an edge, is the thing competitors know about, have probably tried, and still won't do, because it doesn't make sense for their structure. Two moves define WikiDesignCo's alpha, and the market read confirms both are real gaps rather than imagined ones. The first is the verticalized knowledge fabric, which uses the metagraph to pre-encode the standard ontology of a domain, so onboarding a new customer means mapping their data into an existing structure rather than rebuilding from scratch. The enterprise vendors won't productize that for the mid-market, and the agencies can't build it at all. The second is the operator-plus-FDE compression: decomposing the FDE role into a small senior core that designs the playbooks plus a templatized agentic stack that delivers FDE-grade outcomes without staffing an army of $400K engineers, which the product companies won't do because their economics are built on pulling through large software contracts (annual contract value, or ACV) at big accounts. WikiDesignCo's stated three moats (metagraph world-modeling, operator-FDE compression, credit abstraction) are this alpha named from the inside, and the independent market read again picks the same two as the durable ones, the depth and the compression competitors structurally won't replicate.

Wardley evolution and the own-versus-rent call. A Wardley map places each core capability on an axis from genesis (new and custom) to commodity, and the play falls straight out of it. The RAG plumbing (ingestion, chunking, embedding, vector search, the code forked from Archon) sits at product-to-commodity. It's good enough and competed, and reinventing it is the senior-engineer trap, so rent or harvest it and never custom-build it. The metagraph world-model with epistemic provenance, temporal validity windows, and a first-class contradiction engine is genesis-to-custom: novel, differentiating, load-bearing for the user need, and something competitors know about but won't do at this depth, which is the textbook own-and-build capability where the alpha lives. The credit-abstraction pricing layer and the operator-FDE compression model are custom and differentiating, so they're owned too. The bundled twenty-four-tool subscription is composition: assemble commodity and product components behind WikiDesignCo's own promises rather than building any of them. That mapping is the entire don't-build-your-own-database discipline made mechanical: rent the commodity, own the genesis, compose the rest.

Market size and demand signal. There's no clean analyst estimate of the total addressable market (TAM) for FDE-as-a-service plus RAG sold to small businesses, so the read triangulates. Demand for AI implementation is revealed and strong: on top of the FDE postings and hiring figures in the finance read, the role is named one of the fastest-growing in enterprise tech. The small-business base is large and underserved: a 2025 survey found seventy-three percent of small businesses still rely on manual processes for at least three core operations. The willingness to pay at the SMB level already exists, with packaged AI workflows selling at $500-to-$2,000 monthly tiers today, which means the buyer is already spending; they just aren't getting FDE-grade architecture for the money. The category comps from the valuation read set the ceiling. Demand is proven, supply is constrained, and the middle is empty, which is the precise market shape Andy looks for: find where money is already being spent, build something far better, and make it accessible.

7. The build (what this brand needs, where Track R feeds Track P)

WikiDesignCo is the most fully specified of the Looikos brands, so this section is grounded almost entirely in the business brief: its whole architecture is written down.

What it is built from, in two waves. Wave A is live: the Archon-forked RAG core (web crawl, document processing, chunking, embedding, indexing, MCP retrieval) plus the static Library reading surface, the proof-of-platform artifact. Wave B is the production platform and hasn't started. Its stack is decided: Convex as the source of truth, Neo4j plus Graphiti as the temporal metagraph (the validity windows and the contradiction structure), Qdrant for vectors, Typesense for full-text, Inngest for durable functions, Clerk for auth, the Content Compiler agentic stack, the AndyDataBot retrieval surface, and an MCP server for downstream consumers (ContentFactory and the other Constellation brands). The decisive build judgment is the one the repo already made and the value rubric (the scoring sheet Looikos uses to rank brands and capabilities) endorses: the RAG plumbing is a commodity to rent (fork Archon, customize the ten percent that matters), and the metagraph world-model is the genesis capability to own. That's how the open-source repo research feeds every brand build, in miniature: harvest the commodity, build the differentiator.

The hexagonal discipline. The load-bearing rule of the architecture is one core with many interfaces. The graph operations live in a core that never imports a transport, and every interface (HTTP API, MCP server, CLI, web UI, agent interface) is a thin adapter over that core, none of them carrying its own copy of an operation. That's the mechanical defense against what Andy calls the Disconnection, where copies of the same thing drift apart: one authoritative representation per operation, every interface referencing it rather than duplicating it, so a fact or a rule can't drift between the API and the MCP server. The metagraph backend is the canonical example, with Convex, Neo4j, Qdrant, and Typesense all adapters over one core rather than four re-implementations.

The data models. Typed Pydantic models are the intermediate representation (IR), with no ORM. One typed model is the IR across every backend, decomposed ECS-style into entities and components, splitting out to Zod/TypeScript where the frontend needs it. It's the same discipline that the sibling brand Scatter Model productizes, which is why the two sit together in the infrastructure category: WikiDesignCo is the data platform that consumes the IR discipline, Scatter Model is the brand that turns the discipline into a product.

The agent roster the domain needs. Four feature factories, each a set of harnesses plus a gateway harness: ingestion (crawl, process, chunk, embed, index), the Content Compiler (knowledge into in-brand assets), retrieval and AndyDataBot (precise context service), and the metagraph world-model layer (provenance, temporal validity, contradiction). The agentic stack underneath is LangGraph for workflows, PydanticAI and the Claude Agent SDK for the agents inside the nodes, Jinja for prompts, with tools as standalone shareable functions rather than agent-bound methods. Observability is LogFire on every I/O function, evaluation is Hypothesis property tests written by a separate agent from the builder, durability is Inngest.

The medallion asset tiers. The knowledge corpus is tiered from bronze to diamond in the medallion pattern, where each tier is more refined than the one below, with access gated by tier. Raw ingested material is bronze, curated and deduplicated is silver, the validated and provenance-tagged metagraph is gold, and the highest-value distilled expert knowledge is diamond. The credit-denominated access model maps spend tiers onto medallion tiers, which is how the customer's budget selector connects to the depth of knowledge they can reach.

Where the repo research feeds the build. The research into outside open-source repos hasn't started, and Andy will supply the GitHub list. The hooks are nameable in shape even now: WikiDesignCo will want the best harvested patterns for durable agentic execution (the Inngest-style durability layer), for the temporal-graph and GraphRAG layer (whatever the graph and retrieval repos in that research teach), for the agent-harness authoring and cataloging pattern (shared with Symphony AGI and Agent Shipyard), and for the ingestion and scraping layer (shared with Spider Scrape). Once the repo decks exist, the value rubric ranks the combined wish list and the specific capabilities slot in.

7b. The V1 platform build (2026-07-03 re-anchor)

As of 2026-07-03 this section replaces the line in §7 that says Wave B hasn't started. The platform build is in flight under a dedicated team, and Andy's directives that day sharpened the architecture in ways the downstream brands will inherit.

The architecture is a fractal corpus, one pool of assets that every project cuts its own views from. Every asset lands once, in its primitive format, in one central pool: a PDF in the PDF place, a YouTube video with its transcript and comments in the video place, a thread in the social place. The refinery decomposes and enriches each primitive (charts and figures scraped out of PDFs as derived assets, chunking, 3072-dimension embeddings, entity extraction, the full NLP battery), and every enrichment layer is stored as addressable, versioned inventory carrying the model and pipeline that produced it. Above the pool sit the views: a workspace scopes the corpus for a project, and a knowledge base is a curated selection over assets and chunks (fifty of two thousand textbooks, or only the sections of those fifty about one topic, assembled by an agent), each knowledge base carrying its own retrieval indexes and its own MCP connection. The same chunk serves thirty knowledge bases across thirty projects, and an enrichment improvement upgrades all thirty at once. In practice every asset becomes a view over the same source, which is why per-project data infrastructure across the Looikos brands consolidates here instead of fragmenting.

Governance is item-level and load-bearing, because fractal reuse without per-item rules is a breach waiting to happen. Sensitivity, anonymization requirements, usage constraints, and source-client attribution ride on every asset and every chunk as ECS components (entity-component-system, the pattern where each rule is a component attached to the item), and every retrieval path filters on them before anything crosses a workspace boundary. The standard scenario runs like this. A FreelanceBuddy client in oil and gas equipment repair fills a workspace with agent research and five thousand scraped posts; months later a SocialStoryboard client one domain over reuses the shareable portion through a policy-filtered query with an audit trail, and the sensitive material never moves. The same event stream powers the laboratory economics: retrieval events record which chunks earn their keep and for whom, a credit ledger prices every ingestion and enrichment, and the operator tunes intake volume and enrichment depth against measured value, the way a quant desk manages derived-data inventory.

The build decisions on record are these: a separate GCP project cloned from the ContentFactory Cloud Run recipe (projects are GCP's isolation boundary, so a future spin-off is a billing relink rather than a service extraction); the workspace layer on Convex, generalized from the production ContentFactory schema, which retires the Archon-fork direction for the platform while keeping the Archon lesson (one MCP control plane for agents); Gemini and Vertex as the V1 processing baseline, since Gemini is the model that reads video; and a walking-skeleton gate, one workspace proven end to end (create, upload, ingest, search, agent-retrieve on the deployed stack, checked with the three-proof quality check) before the first-wave roster of twenty workspaces is seeded from the manifest. The whole build follows one staging rule: the platform is the environment, humans and command-line agents use it through the same interfaces, and resident agents that live on the platform (Hermes-class) arrive later, after the low-level processing. The first real workload is the world-model business brief, one per workspace, a first-class asset type from the first schema push, because the line between a workspace and its world model is the platform's founding contract.

7c. The four faces and the memory layer (2026-07-04 unified-vision re-anchor)

Research on 2026-07-04 filled in the whole-platform picture that §7b began, so the deck carries the full vision the downstream brands inherit. WikiDesignCo is one platform with four faces, the same thing seen from four sides, and §7b covered the first two in build terms.

The first face is the central knowledge platform: one central pool of assets typed by primitive format, with workspaces and first-class knowledge bases as fractal views cut from it, the mechanism §7b already detailed. The second is the client-powering engine: a workspace per company and client, a world-model business brief of roughly fifty-five aspects as the first workload, and a domain authority map as the workspace home view. The map shows, for each element of a domain, the top entities, how much source material backs each, how fresh it is, and where the thin spots are, so the operator sees at a glance where they're already an authority and where to write or ingest more. Publishing into a gap visibly moves the map, which is the game feel the laboratory loop promises, made concrete on a Tuesday.

The third face is new to this deck, and it's load-bearing: WikiDesignCo is the unified memory layer for every Looikos brand, holding CRM threads, agent conversation threads, and customer-service threads alongside documents, all landing as assets in workspaces. Experiences consolidate into episodes that carry a summary plus pointers back to the exact turns; episodes accumulate until an assessment produces improvement proposals; every proposal resolves to a permanent decision record, greenlit or archived-with-reasoning, so a later agent can query why something was decided and when, and never re-proposes a rejected idea blind. Many agents share one memory, which makes the platform fractal in memory as well as in content. The fourth face is the metagraph itself: every fact carries provenance, confidence, and when-it-was-true automatically, as a habit, while the machinery (reified statements, bi-temporal validity, a contradiction engine) stays abstracted behind plain words in the interface, Sources and Coverage and Freshness and Where-this-came-from, never graph-theory vocabulary and never the Goertzel or quantum framing that belongs only in the private lineage docs.

The file-format decision on record is that the Pydantic intermediate representation emits Open Knowledge Format (OKF) bundles as its portable file form, one that lives in git and that people and agents can both read. OKF was adopted first-class because its flat typed frontmatter with one required key is the exact shape of a thin Pydantic node, so serializing to it is a short path rather than a translation project. Any agent, ours or a client's or a third party's, can read OKF without custom glue, while the claims layer stays where the reasoning happens. One intermediate representation feeds several serializations (Graphiti episodes, vector records, OKF files). The cost of the bet is near zero because a bundle is a directory of markdown in version control, so if the standard stalls the platform still holds readable typed files.

8. Priority read (feeds the value rubric)

WikiDesignCo is foundational substrate, the highest-leverage brand in the ecosystem to stand up, because the other infrastructure brands and the agencies inherit its metagraph, its infrastructure, and the trust it establishes. On the graph of which brands depend on which, it's a foundation node: many leaves depend on it, so it's sequenced first regardless of how any one leaf scores. The single most important thing this read hands the value rubric is a rule: never rank a brand that depends on WikiDesignCo's metagraph above WikiDesignCo itself.

Readiness is split by wave, and the wave split is the gate. Wave A (the RAG core and the Library) is live and proven; the ContentFactory live instance is the customer-facing receipt that the model converts. Wave B (the production metagraph platform) hasn't started, and Wave B is what the downstream brands actually inherit, so the priority read has to keep the live receipt separate from the unbuilt substrate rather than treating the brand as uniformly ready.

The first-pass tiering goes capability by capability rather than treating the brand as one block, because the unit the rubric prioritizes is the capability:

Now (build and own): the metagraph world-model with epistemic provenance, temporal validity, and the contradiction engine. It's genesis-stage, load-bearing for the user need, the alpha competitors won't replicate at depth, and the thing every downstream brand reads. It's the highest-leverage build in the ecosystem, and because it shapes many future decisions it's judged the way Powell's value-function approximation (VFA) judges substrate: scored on its discounted future rather than its immediate fit.
Now (compose and ship): the Wave-A RAG core and the Library, already live, the proof artifact that earns the trust the rest inherits.
Next (gated on the Now substrate): the credit-abstraction pricing UI and the Content Compiler agentic stack. Both are differentiating and both depend on the Wave-B metagraph being real, so they're attractive but blocked until the foundational promise is kept. They're judged by Powell's cost-function approximation (CFA): the direction is set and the parameters need working out.
Leave (rent, never custom-build): the RAG plumbing itself (ingestion, embedding, vector search). It's a commodity, and forking Archon already rented it, which was the right call. Reinventing it is the senior-engineer trap the whole ecosystem is disciplined against.
Watch: the specific open-source harvest targets from the repo research for the durability, graph, and ingestion layers, which become rankable only when Andy's repo list lands, and that's the trigger to revisit them.

The seven-sins check runs against this read too. Pride or look-ahead: the read scores Wave B as unbuilt and the metagraph as a bet, not as if it already shipped, so the present is scored and the bet is flagged. Envy or survivorship: the failure cases are in the deck (the six-week flaming pile, the failed pilots) rather than only the dealership win. Gluttony or overfitting: the enthusiasm is capped to the one live receipt, not inflated by the twenty-four-tool feature count. Sloth or transaction-cost: the Wave-B build friction is named as the gate, not rounded away. Wrath or regime-blindness: the read assumes the 2026 FDE-scarcity regime and the token-price-falling trajectory, both stated. Lust or capacity delusion: WikiDesignCo is one foundational build, not an attempt to ship all four factories at once. Greed or fat-tail: the tail risk is the metagraph build proving harder or slower than the differentiation justifies, which is why it goes through value-function approximation and gets the discounted-future treatment rather than a rule-based adopt. One dependency reaches past this deck: WikiDesignCo's Wave B is the keystone the infrastructure and agency roadmap arches over, so its sequencing decision orders much of the rest.