What we have built
Every AI consultancy shows client logos. We would rather show systems we own, because we can tell you how they work, what broke, and what they cost to run — none of which we could do with someone else’s project.
A lead-capture agent for medical clinics
A chatbot product for clinics and medical tourism operators. It answers enquiries in the visitor’s language, asks the intake questions a coordinator would ask, and produces a structured lead rather than a transcript. It runs on WhatsApp as well as the website, and a human can take over an in-progress conversation at any point — which turned out to be the feature that made clinics willing to switch it on.
What we learned: handover is the whole product. An assistant that cannot be interrupted by a person is a liability, and the clinics that trusted the system were the ones that had watched themselves take over mid-conversation.
Built with: a self-hosted open model, a message gateway, structured output validation, per-clinic knowledge bases.
An SEO platform with its own crawler and link graph
A full analytics platform: a polite crawler that honours robots directives and crawl delays, around twenty-five technical rules scored into a site health figure, a domain authority metric derived from a web graph of roughly seven billion links, and a local model that turns the raw findings into a prioritised list of things to do.
What we learned: the model must never produce a number. Every figure in the output comes from our own measurements and is passed to the model as fact; the model only decides ordering and explains the fix. That single rule removes the invented-statistic failure mode that makes most AI reporting unusable.
Built with: a columnar database holding billions of rows, a distributed crawl, self-hosted models for the analysis layer.
An eleven-language content migration
Around five thousand pages of a travel site translated into eleven languages using models we host, with a glossary that locks product and place names, and — the part that matters — real localised URLs with reciprocal hreflang rather than a browser-side language switcher.
What we learned: the translation is the easy half. The URL structure, the canonical handling and the sitemap logic determine whether search engines ever see the work at all. We have watched a seven-language switcher produce exactly zero indexable pages, on our own site.
Built with: batched local inference, a review queue for commercially important pages, static generation.
Why we show our own work
Because it is verifiable and because it is where the honest lessons are. A case study written about a client is a marketing document approved by that client. A case study about a system you run yourself can include the parts that went wrong, which are the only parts that tell you anything about how a supplier will behave when your project goes wrong.
Client references are available under NDA once a conversation is serious. We do not publish client names as decoration.
Tell us what your team keeps not getting to.
A 30-minute call. We tell you whether an agent is the right tool for it, what it would take to build, and what it would cost to run — before anyone signs anything.