Cobaltica — AI Engineering Studio
Our products

What we build for ourselves

Alongside client work we develop our own products. They are not a portfolio: they are where we try out first what we later propose to clients — AI agent architectures, verified generation, multilingual content, white-label delivery.

ludica.ai

Escape rooms and multi-stop narrative trails, anywhere.

The platform museums, schools, companies and public bodies use to build playable experiences in a real place: scan a QR code and play in the browser, with nothing to install. Compass and GPS coordinates outdoors, navigation puzzles indoors, a live leaderboard and narrated audio.

The part that interests us most is the editor: AI generates stories, puzzles, voice and translations from the real context of the trail, and verifies that every puzzle is actually solvable before it reaches players. The author stays the author — the machine takes away the mechanical work.

Go to ludica.ai
ludica.ai: the compass pointing to the next stop, the adventure cover and the opening story
What it proves
AI generation that is verified, not just generated
Architecture
Multi-tenant white-label, offline-first
Who it is for
Museums, schools, companies, public bodies
Content
Natively multilingual, with narrated audio

Sibillia

Cultural heritage that speaks to the people visiting it.

The platform museums, historic towns and whole territories use to turn their heritage into an interactive narrative experience. Visitors open it in the browser, from a QR code or straight along the route, and can listen to stories, explore works and places, and put questions to an AI assistant that answers from the content and the sources the institution has chosen.

The same visit can change language, depth and tone of voice: from an explanation written for adults to a story made for children and families. Text, audio and conversation live inside the same experience, while every museum or territory keeps its own identity, its own content and a dedicated environment. AI does not replace the knowledge built by historians, curators and institutions: it makes it reachable in different forms, at the moment the visitor needs it.

Go to sibillia.ai
Sibillia on three phones: the Maria Montessori entry with its narration playing, the visit trails of a church in Cascia, and the lunette of the Sibyls in the Nobile Collegio del Cambio, each with the button to talk to the AI assistant
What it proves
Conversational AI grounded in verified cultural content
Architecture
Multi-tenant, white-label, RAG and AI agents
Who it is for
Museums, cultural institutions, municipalities and territories
Content
Multilingual, narrated audio, trails for different audiences

foodmanager.info

The food inventory behind hospitality schools' training kitchens.

The platform hospitality schools use to run the food stores behind their practical lessons. Teachers file requests from the browser, the office aggregates them and checks them against stock, storekeepers book goods in and hand them out: what used to be a round of forms and phone calls becomes a single flow, with the status of every request always visible.

From the supplier order to delivery in the kitchen every step stays on the record, with estimated spend alongside actual. It runs in the cloud with nothing to install: a browser is enough.

Go to foodmanager.info
foodmanager.info, the school office dashboard: requests to approve and to close, orders, estimate against actual, the school budget split into spent, committed and available, and spending by supplier
What it proves
A management system in production, three roles on a single flow
Architecture
Multi-user cloud application, nothing to install
Who it is for
Hospitality schools: teachers, office, storeroom
Data
Requests, stock levels, estimated against actual spend

Thread

In internal use

The thread that ties meetings, decisions, tasks and documents together.

The platform we work on every day. It connects clients, projects, meetings, decisions, tasks and documents: a recorded meeting becomes a transcript with who said what, a summary and tasks ready to assign; quotes are drafted on letterhead from the project's context; a single search runs across meetings, mail, documents and calendars.

AI is asked for by capability, not by vendor: transcribing, summarising, searching are requests that configuration routes to the right model, with a fallback when one does not answer. No AI output takes effect without a person confirming it, and everything runs on our own servers.

Thread, the Today page: meeting and task indicators, the queue of things to do — a summary to confirm, a quote due today, tasks to assign — and beside it the day's agenda and the AI jobs running
What it proves
Multi-vendor AI requested by capability, always under human confirmation
Architecture
Modular monolith in .NET and Angular, self-hosted
Who it is for
Our own studio, before anyone else
Data
Meetings, transcripts, tasks, documents and mail

Why we build our own products

Because a supplier who has never taken a product of their own into production does not know what it costs to keep one running. Our products pay their own bills, have real users, and force us to solve the same problems you will face: model costs, quality over time, multilingual content, updates without taking the service down.

Tell us about your project

Describe the process you would like to automate. In half an hour we will tell you whether an agent makes sense — and if it does not, we will tell you that too.

Where we are
Via La Louvière, 2
06034 Foligno (PG) — Italia

Up to 5 files, no more than 15.1 MB in total. PDF, Word, ODT, RTF, text or images.

May we measure visits? We would use Google Analytics to know which pages get read. Nothing else: no advertising profile, no data handed to advertisers. Until you choose, measurement stays off. What this involves