Design, Build & Brand Technology Studio

Good engineering decisions come from evidence, not vibes. That's why we run and publish open-source research for developers of LLM applications: a growing series of pre-registered studies, more than 31,788 scored model runs and counting, corrections included.

The flagship is Barkup Bench, our open benchmark measuring how language models read and edit structured documents. Every study is pre-registered before a single model is called: forty-two studies so far, across five models, trees from 5 to 1,000 nodes, and editing sessions up to 36 edits long. Everything the benchmark validates ships in our open-source barkup library, and everything it refutes gets published anyway.

We believe great strategy is the process of asking tough questions. Great design is the process of making tough decisions. Great implementation is the process of making tough compromises.

Easy answers are everywhere. The turnkey proprietary platform looks easy until you factor in the true costs and limitations; a customized open-source solution is usually more flexible, less constrained, and a better price. The LLM-generated brochure looks easy until you try to tweak a sentence, and it never quite matches your brand guidelines anyway. In the LLM age, easy is the easy part. The hard part is the guardrails: custom applications that keep generated collateral on-brand and on-message, and code integrity rules and tooling that keep quickly generated code from becoming a tech-debt nightmare.

Open source is an open, collaborative, and community-driven approach to software development. It promotes transparency, innovation, and cost-effectiveness. Whatever your technology challenge, there is probably an open-source solution that is the best choice to solve it.

That belief shows up in the numbers: two-thirds of the technologies we work with are open source, and this very website runs on an open stack of SvelteKit, UnoCSS, Vite, and Bun (the full inventory lives on our Built With page). It also shows up when it costs something: when Pimcore, a platform we implemented for more than a decade, moved to a source-available license, we backed OpenDXP, the community fork keeping the open codebase alive.

It has never been easier for a business to get swept up in AI hype, and never harder to tell the real opportunities from the noise. We help businesses adapt their marketing technology and sales tools to the LLM era with a customized, research-driven approach: test the claims, filter out the fiction, and invest only in what the evidence supports.

We know how convincing fiction can be, because our own lab once produced some. An early study in our benchmark series published a dramatic 33-point effect that turned out to be a bug in a popular SDK, so we retracted the finding in public and rebuilt the study. That is the discipline we bring to your stack: the same claims-testing we apply to our own headlines, applied to every vendor pitch, plugin, and platform that wants your budget.

Your brand is not what you say about yourself. It is what customers experience: every page, every search result, every checkout, every reply. The way customers interact with you becomes your brand. That is why we practice brand and digital as one discipline, from positioning and tone of voice to the interfaces that carry them.

And increasingly, the one experiencing your brand is not even human. AI assistants now read your website and describe your business to customers in their own words, and generated collateral carries your name whether or not it sounds like you. That is why we built Replicator, our brand operating system, where a language model designs and drafts inside brand-approved rails: positioning, voice, and guardrails encoded so the brand holds even when a machine is doing the talking.