Protect
the Alpha

ShurIQ is a modular framework of roughly a dozen components wired together by a custom orchestration harness, built deliberately against the one-vendor-per-tier assumption. It rents the commodity layers and owns the layers that encode the agency process. The rented layers are model weights, hosting, the vector store, and embeddings. The owned layers are the SBPI rubric, the grammar canon, the workflow orchestration, and the portability library. Each layer carries three facts: what is proprietary, what is rented by choice, and how mature it is. Proprietary layers can still be early, and deliberate vendor dependencies can be fully live today.

Five components are built and running in production right now: the multi-agent generation pipeline, the SBPI knowledge graph and rubric, the shuriq-kg portability library, the deterministic grammar gate, and Report Studio. They have shipped deployed client reports, they have measured runs, and they compound nightly. Most AI startups at this stage cannot claim a single one of these as a standing system.

Legend. Owner: us (proprietary), vendor (paid third party), OSS (open-source self-hosted), hybrid (our code over rented or OSS pieces). Horizon: owned-now = live and running today; near-term = partially built or planned next; long-term = design-stage; never = rented permanently by design, not a build target. State: BUILT/MEASURED versus PLANNED/TARGET, called out per row.

Proprietary and live now: the SBPI rubric and per-vertical ontology, the report grammar canon with the deterministic gate and banned-terms list, the shuriq-kg portability library, the workflow.js orchestration and harness, Scout's negative-space method, the Report Studio app code, the IP-firewall export filter, and the two-tier protocol. These encode the agency process. None of them is observable from a published brief, so a competitor cannot reconstruct them by reading a finished output.

Rented by design: foundation-model weights, OpenAI embeddings, Cloudflare hosting, and licensed data. CarbonArc is the licensed-data example: access to transaction velocity and zip-level economic signals, a dataset no single client could assemble alone. Integrated into the framework, that data flows through the same owned judgment layer and lands in the customer's own graph: licensed signal in, owned intelligence out. Proprietary value lives above the model layer, so renting these is the strategy. The whole stack reads one injected model parameter, CLAUDE_MODEL, defaulting to claude-opus-4-8 in both the workflow pipeline and Report Studio. Swapping providers is a configuration change. Portability is already proven at storage: two triplestore adapters, rdflib and Oxigraph, run identical SPARQL over the same data, and 1,501 triples returned identical results in both on 2026-07-28. A live second-provider generation path is the next milestone.

Revenue today is engagement-led: founder-led sales and roughly 60 analyst hours per engagement. This is the deliberate sequence: prove the scoring method against real client work, compound it through the running system, then package it. The SBPI self-score of 40 reflects early distribution.

The website output, the hosting, the data storage, the report layouts. Weeks of work for a funded team, and we treat none of it as an advantage.

1. The trained scoring. The categories and the point scale are public craft. The judgment about what a piece of evidence is worth took two years of real client work to train, and you cannot recover it by reading the finished scores. We tested this on ourselves: an outside team studying our reports could reproduce how they look almost perfectly, 90 to 100 percent, and the judgment underneath at only 5 percent.

2. The writing standards. Any AI writes fluent text. Ours carries the accumulated list of everything real clients rejected, corrected, and approved across 100+ shipped reports. That list exists only because the work shipped.

3. The knowledge that builds itself. Every report, meeting and decision adds to a knowledge base the client owns: 119,054 connected facts as of 2026-07-29, growing every night. A competitor cannot buy it or scrape it, because it is generated by the work itself.

A complete report costs about 45 dollars to produce as of 2026-07-29, down from about 115 dollars on 2026-06-01. Nobody else in our field publishes that number.

Specialized models trained for each client's industry, running inside the client's own walls for the customers whose security requires it. The system already treats every AI model as a replaceable part, so these drop in without a rebuild.

Shur Creative Partners is the other half of the wheel. The agency shipped about one million dollars of strategy, brand and content work in the past year. ShurIQ makes that work faster and better, and every project teaches the system more. The work feeds the intelligence, and the intelligence improves the work.

The model is simple: enterprise customers times average contract value, plus services revenue. The contract assumption comes from live pilot pricing of $7,500 per month per brand plus a creative retainer, which puts the working average contract at $150,000 per year.

Every number in this table is a model built on stated assumptions, never a measurement. For scale: a company on track for a $1B valuation typically shows $40M to $60M of annual recurring revenue at year five.

Sixteen years of research and thirty years of professional judgment across the agency.

Jonny Dubowsky, co-founder. Sixteen years designing AI agents across multiple industries, all in service of one idea: compounding and augmenting collective intelligence. Lectures at NYU ITP, Syracuse University, and Drury University, with ongoing workshops alongside leading researchers. Architect of the judgment engine ShurIQ runs on.

KG Hagedorn, GM, ShurIQ. Serial entrepreneur who has scaled companies past $100M in revenue in under five years. EMBA, NYU Stern; Program for Leadership Development, Harvard Business School.