Source: Unite.AI
Guglielmo Faglioni, Co-Founder and Chief of Science & Technology at Callimacus, is a technology entrepreneur and software engineer with more than a decade of experience building digital products and applications. Alongside his role at Callimacus, he has served as CEO of Gorilla Technologies since 2016, a software development company focused on mobile applications, websites, and e-commerce solutions. Earlier in his career, Faglioni founded GuiDev, where he worked as a lead iOS software engineer. At Callimacus, he helps lead the scientific and technical development of the company’s approach to AI-driven digital experiences.
Callimacus is an AI platform developed by Milan-based Solomei AI that replaces the traditional concept of static web pages with digital experiences assembled dynamically for each visitor. Its architecture uses an orchestra of specialized AI agents to interpret signals such as questions, clicks, and scrolling behavior, determine user intent, and compose relevant interface elements in real time. The platform can connect with systems including Shopify, Salesforce, content management systems, and APIs while providing developers with SDKs for integrating its intent and interface-generation capabilities into existing front ends. Callimacus has also been deployed to power AI-driven digital experiences for luxury fashion company Brunello Cucinelli.
Before co-founding Callimacus, you spent years building software as an iOS engineer and later founded and led Gorilla Technologies. What experiences from that journey convinced you that the traditional, page-based architecture of the web was ready to be fundamentally rethought for the AI era?
Building iOS apps taught me something that took years to fully appreciate: every screen is designed in advance, and every user sees the same one. You draw the app, you ship it, and then you watch millions of different people push themselves through an identical flow. You can segment and A/B test, but you are still choosing among versions somebody drew beforehand. The web inherited the same model from print: the page as a fixed container. Once models became capable of understanding intent and making structured decisions quickly enough, it became possible to question the container itself instead of optimising what goes inside it.
Callimacus replaces predefined website journeys with an “orchestra” of AI agents that interprets visitor intent and composes an interface in real time. Can you walk us through what is happening technically from the moment a visitor arrives to the moment that personalized experience appears?
Callimacus is built around a pageless architecture. Instead of organising content into predefined webpages, our technology separates content from the page itself. Products, editorial content, media and interface components exist independently, so they can be recombined dynamically rather than being locked into fixed containers.
As the visitor begins interacting, Callimacus analyses multidimensional signals such as clicks, navigation patterns and prompt-based conversations. The Intent Agent uses those signals to understand the visitor’s context, interests, expectations, behaviour and what they are trying to achieve at that particular moment. From there, the Intent Agent activates the most appropriate “receptor” for that motivation. For example, somebody looking for inspiration may follow a different interaction path from somebody showing a clear intention to purchase. The system then selects the most relevant content and interface components and composes the experience within milliseconds. That process does not happen only once. As the visitor interacts and their intent changes, the pageless experience can continue to recompose around them, rather than forcing them through a journey that was designed in advance.
Personalization has existed on the web for years through recommendation engines, segmentation, and A/B testing. What makes intent-driven generative interfaces fundamentally different from those earlier approaches?
For us, the fundamental difference is between personalising from the past and understanding what somebody needs now. Most existing personalisation is based on things such as CRM profiles, cookies and historical behaviour. That can tell you something about what people have done before, but intent is much more immediate and can change during a single session.
Callimacus is designed to understand that intent from real-time signals and personalise at the individual level. It is also not limited to changing a recommendation or one element of a page. The content, interface components and interaction path can all be composed around that person’s current intent and then recomposed as it evolves. So the shift is from choosing among predefined experiences based largely on historical information to generating the experience itself around what the individual is trying to achieve in that moment. Historical data can provide context, but real-time intent determines how the experience adapts.
If every visitor can effectively receive a different website, how do you prevent personalization from becoming unpredictability? How do you preserve a brand’s identity, design language, and business rules while still giving AI meaningful freedom to adapt the experience?
Personalisation does not mean giving the AI unlimited freedom. Brand control is built into the architecture, so the experience is always generated within the brand’s approved rules and assets. Every interface element comes from a registry of components the brand has approved, already designed to its visual language. The model produces a structured description of what to show and how to arrange it, validated against a strict schema before anything renders. In other words, the model decides the composition, while the brand’s design system decides what any composition can look like. Business rules such as pricing, availability and regional assortment are enforced outside the model as hard constraints.
Control also means visibility. Through the backoffice, business teams can see how the website is behaving: which experiences are being composed, for which intents, and how visitors respond. The platform also diagnoses itself, and does so proactively. It points out where it might need improvement, tries new configurations, tests them against previous user interactions and evaluates whether they work. Nothing goes live until the brand approves it, so the AI proposes improvements backed by evidence and the brand keeps the final decision.
That is important because one of the problems we are trying to solve is that traditional websites have made many brands look and behave increasingly alike online. AI should give a brand greater ability to express its identity around the individual visitor, while the brand continues to define the boundaries within which that experience is created.
Real-time generative interfaces introduce new reliability challenges because AI is no longer just generating an answer; it is helping determine what the user actually sees and how they navigate. How do you approach hallucinations, incorrect intent inference, safety, and verification in that environment?
We start by limiting what the AI can get wrong. The model decides how the experience is composed, but the facts inside it come from the brand’s systems of record. Intent inference is treated as a probability. When the system is uncertain, it composes a broader, more exploratory experience rather than committing to one interpretation. Because the experience keeps recomposing as the visitor interacts, each new signal corrects the previous guess.
Reliability is also a matter of architecture. The layer that delivers the experience is isolated from the AI layer and held to a higher availability target. If a model is slow or unavailable, visitors still get a complete, on-brand experience built from deterministic defaults.
The Brunello Cucinelli AI boutique is one of the first major real-world deployments of this model. What have you learned from observing how customers interact with an AI-composed commerce experience that you did not anticipate when building Callimacus?
One of the clearest signals has been that people are willing to engage more deeply with an experience that feels more relevant to them. We have seen users spend around twice as much active time on the website, alongside a significant uplift in both add-to-cart and conversion rates.
For us, that has been an important early validation that moving away from fixed, one-size-fits-all journeys can have a meaningful impact not only on how distinctive the experience feels, but also on engagement and commercial performance.
Dynamic websites also challenge many assumptions behind search engine optimization, analytics, accessibility, and testing. How do technologies built around relatively deterministic web pages need to evolve when interfaces and user journeys are generated dynamically?
Most web tooling assumes the page is the unit of everything: you index a page, measure a page, test a page. In a pageless architecture that unit shifts to the content itself, and each discipline has to follow. For search, every piece of content exists independently of how it’s composed, so it can have its own clean, language-aware URL and appear in an automatically generated sitemap across every locale. The next step goes further: the platform detects the navigation paths visitors actually use most and exposes them to search engines, and it serves trusted first-party content directly to AI agents and generative search engines. Discoverability then reflects real demand instead of a sitemap designed in advance.
Analytics moves from pageviews to intent. Because the system interprets what each visitor is trying to achieve, including anonymous visitors that traditional analytics miss, brands can analyse behaviour at that level and export those insights into their CRM or CDP. Teams can also query this conversationally and build their own dashboards.
Testing changes the most. You can’t click through every possible journey when every journey is unique, so we backtest the agents against real past interactions and monitor them continuously in production. Accessibility follows the same principle: it’s built into each component, so every composition inherits it.
For real-time AI experiences to feel natural, inference needs to happen with extremely low latency. What are the hardest engineering challenges involved in putting AI directly in the critical path of a website, particularly around model selection, caching, orchestration, cost, and fallbacks?
For inference, we currently run on Groq’s LPUs, which give us very low latency, and we orchestrate multiple models, routing each task to the model best suited to it. We test new models as they come out, looking for the best fit for each task in terms of quality, cost and latency. Because the architecture is model-agnostic, we can swap in a new model or inference provider quickly whenever it’s worth it. Depth also adapts to the request: a simple query gets a fast answer, and only nuanced requests trigger deeper analysis. This keeps cost under control, since the heaviest reasoning is used only when it changes the outcome.
Fallbacks are part of the architecture from the start. The layer that delivers the experience is isolated from AI enrichment, so if a model slows down or fails, the visitor still receives a complete experience.
Salesforce recently invested in Callimacus as the company looks to expand its engineering, AI research, and commercial operations. What did Salesforce see in the technology, and how could connecting generative interfaces with enterprise CRM and commerce data change what websites are capable of doing?
What Salesforce recognised is that Callimacus is not simply adding AI to an existing website. The platform is rethinking the digital experience itself, combining conversational AI, enterprise context and real-time personalisation to create new ways for companies to engage with customers. Today, CRM and commerce systems contain a huge amount of context about customers, products and interactions, but the website experience is still largely static. Connecting that context with generative interfaces creates the possibility for websites to become much more responsive. Not just displaying information, but adapting content, products and interactions in real time around what a visitor is trying to achieve. For us, that is one of the most important opportunities ahead. The website can become an intelligent interface between a company’s existing systems and the individual customer, rather than simply a fixed presentation layer.
As AI agents increasingly browse, research, compare products, and potentially transact on behalf of people, could the bigger shift eventually extend beyond websites adapting to human visitors? How do you envision the web changing when brands need to serve both humans and autonomous AI agents at the same time?
As more discovery and shopping move into AI assistants, brands risk losing control over how they are represented. A rich brand experience can easily be reduced to a product image, a short description and a price inside a third-party interface. That creates a new challenge for brands: how do you stay useful to AI agents without giving up the identity, storytelling and direct relationship you have built with customers? Adding a chatbot to a website doesn’t solve this. A conversation window can answer questions, but it can’t carry the full expression of a brand: its visual language, its craftsmanship, the way it tells stories. Brands will always need a place they control, where they can express themselves on their own terms. We believe websites will remain that place, even as more of the customer journey becomes mediated by AI, and they will need to adapt to every visitor, human or agent.
Thank you for the great interview, readers who wish to learn more should visit Callimacus.