The Material, Not the Feature

There is a scene that has been repeating itself, for a couple of years now, in meeting rooms across Italy. A vendor plugs in a laptop, opens a demo, and shows an assistant that answers any question about the company's documents with effortless fluency. The tone is right, the answer arrives fast, someone takes notes. Six months later, that assistant is nowhere to be found in the company's actual processes: the demo ran on ten well-chosen documents, reality contained forty thousand badly chosen ones, and nobody had decided who answers when the system gets it wrong.

The distance between that demo and a system that works every day is not closed by a larger model. It is closed by structure: the code that surrounds the model, feeds it, supervises it, and records its decisions. This is the conviction behind AYXZA, an Italian studio of AI software systems: building software in which intelligence is not an added feature, but the very material the structure is made of.

Products, not projects

The studio's founding choice is a single one, and it is worth stating upfront: building products, not bespoke projects. The difference is not commercial; it is an engineering one. A bespoke project is born for a single company, inherits its constraints, and once delivered leaves behind an orphaned system nobody wants to maintain. A product, by contrast, is forced to generalise: it must work for different clients, survive model upgrades, and degrade predictably when something breaks. It is a harder discipline, and precisely for that reason it produces more solid systems.

There is a second choice, less visible but equally structural: every product in the catalogue rests on the same infrastructure. Not nine independent codebases, but a common ground layer on which each system builds its own specifics. When that layer improves, all of them improve together.

Five domains, nine products

The studio's catalogue currently counts nine products, organised into five domains that are not internal divisions but angles of the same practice.

The first domain is infrastructure. AYXZA Relay is a gateway library that unifies access, routing, fallback, caching and knowledge bases across OpenAI, Anthropic, Gemini, Vertex AI and any compatible endpoint: five providers, one Python interface. It is the ground layer just described, the one that lets you swap models without rewriting the system, keep costs contained, and avoid tying your fortunes to a single vendor.

The second domain concerns documents and knowledge, and it is the most populated. Vdoc is a document management system that collects company documents from four channels, interprets them with a proprietary AI engine, and files them with the correct metadata and permissions: it reads an invoice, a contract or a receipt and classifies it on its own. BA.IA. addresses a problem familiar to every Italian business: dozens of public grants and incentives are published every week, and almost none are relevant; the system gathers them from multiple sources, evaluates each one against the company's profile, and delivers only the pertinent ones, with a score and the reasoning behind it. forge_qa, finally, is AI-assisted technical support for software vendors: the intelligence classifies incoming requests, searches the product manuals and drafts the answer; the operator confirms with a single gesture when the answer is already good.

The third domain is that of multi-perspective advisory systems, perhaps the studio's most distinctive idea: virtual consultants that read the same problem from different angles and debate each other, because the best decisions are born from confrontation, not from a chorus. Strategic advisory for entrepreneurs accompanies the decisions that change a company, from generational handover to holding reorganisation to corporate evolution, in the phase of life when the great strategic questions of an Italian SME all arrive at once. BO.I.A. applies the same principle to markets: several specialised AI analysts examine the same listed securities from opposing viewpoints, argue, contradict one another, and an editor-in-chief brings order and prepares the morning briefing.

The fourth domain is operational platforms. Joshua Cloud Infrastructure is a multi-tenant SaaS platform built on a four-level hierarchy, with native data isolation at the database level and an integrated AI gateway that brings language models inside every operational flow, not beside it. AYXZA Tutor is the adaptive learning platform that descends from it: whoever creates a training project uploads the manuals, the AI generates the learning path, calibrates quizzes where they are needed, and rephrases explanations in the style that works for each learner. One single system for companies, professionals and students.

The fifth domain, creative environments and simulation, contains the most lateral and perhaps the most ambitious product. Project Oreste is a space where anyone, not only developers, can invent a board game, write the rules once, and have them enforced automatically: a three-level editor, a secure rules engine identical in the browser and on the server, and a pipeline that reaches all the way to the physical printed game.

Nine different products, then, but a single practice: in each of them, models decide, perceive, write or regulate inside the software's structure, never as an ornament resting on top of it.

Half model, half code

Behind the catalogue there is a method, which can be summarised in two convictions.

The first: a serious AI system is half model and half classical code. Request routing, versioning, automated evaluations and final checks matter as much as the quality of the prompt or of the network. Whoever treats the model as a magic box to be queried is building a demo, not a system; whoever treats it as a component to be engineered, with clear contracts on input and output, is building something that can stay in production.

The second: no magic. Every decision taken by one of the studio's systems leaves a trace that is readable, reproducible and revocable. "Magic" is the way systems break silently, and systems that break silently are the most expensive kind there is, because the damage is discovered only once it has matured. Better to break loudly, and get back up documented.

From these two convictions follows the studio's commercial posture, which serves two distinct audiences. For software houses and IT resellers, AYXZA is the specialist AI department that is not worth building in-house: design and components in white-label, inside products that carry the reseller's own brand. For companies with real problems, the studio builds directly, with the same product discipline.

A note of caution

There is a serious objection to the very idea of building AI products today, and it deserves to be taken seriously. It goes like this: models improve on a quarterly cadence, and every new generation absorbs capabilities that until yesterday required dedicated software. Whoever builds on top of the models, the objection says, builds on sand: today's product is tomorrow's native feature.

The objection carries a share of truth it would be dishonest to deny. Many products born in recent years are thin layers over an API, and a thin layer evaporates at the underlying provider's next release. Those who invested in such bets have often lost.

Yet it is precisely this dynamic that justifies the opposite approach. The value that models do not absorb lies in the structure: in the company's data and the permissions that govern it, in the operational flows the system is grafted into, in the infrastructure that allows one model to be replaced by the next without rewriting anything. A system built this way does not fear the progress of models: it collects it. When the model improves, the system improves with it, because the model is a replaceable component and not the product itself. It is the difference between building on sand and building on the shore, knowing exactly where the tide will reach.

What remains

A studio is judged by what remains in operation after the presentation meeting is over. AYXZA's wager is that AI software deserves the same seriousness demanded of any other software: versions, checks, traces, accountability. Not demos that shine once, but systems that work every day, in production, for real clients.

The full catalogue, with a dedicated page for each of the nine products, is at ayxza.com. To talk about it: ping@ayxza.com.