Case study: Carte

Carte is a platform designed to publish and work with poster collections. Its first collection is the Carlos Velasco Collection, with more than 7,500 posters covering propaganda, health campaigns, culture, tourism, and political activism: a primary source for researching the visual culture of the last century, which, until now, was kept in an archive with limited external access.

Bringing the archive into a platform that would be useful for three types of users with very different needs:

• Researchers and universities, who need accuracy, the ability to combine search criteria, and to verify the provenance of each piece of data.
• Teachers and mediation teams, who often arrive with a session to prepare and very little time to put it together.
• The general public interested in visual culture, who search without technical vocabulary and may give up if the first search returns no results.

The project came with two firm requirements: automatically generated data could never alter the original cataloguing information, and the original files had to remain in private storage.

After reviewing the archive and working with the team responsible for the collection, we proposed a platform built around a specific architectural decision: to store the original cataloguing data, AI-suggested data, and human-validated data separately. It is human-in-the-loop AI built into the data model: AI adds information, and people decide.

This means that for any field in a record, it is possible to know where the information comes from: the original cataloguing, a model that inferred it from the image, or a professional who validated it. Without this separation, automatically generated enrichment would have become mixed with the original cataloguing data, reducing the value of the collection.

On this foundation, we developed two different environments: a catalogue for finding items and educational toolkits for preparing complete activities. They share the same collection and search engine, but each is designed around its specific purpose.

The Catalogue and Its Filters

Each poster retains its historical and archival metadata: title, authorship, date, location, sector, historical period, and visual and thematic classifications. On top of this, we developed a system of combinable filters by period and date, authorship, location, sector or theme, artistic style, iconography, visual aesthetics, target audience, message, and historical context.

The iconography, aesthetics, and target audience filters are what truly broaden access to the collection: they make it possible to search for images based on their content, what they represent, or their visual treatment, without needing to know a title or an author. Healthcare posters featuring children, which previously could not be easily grouped together, can now be found with a single search.

Presenting and Comparing Items

The toolkits include a presentation mode for displaying posters to a group, with sequential navigation through the approved selection, and a comparison mode for displaying several items at the same time. From the guide, users can jump directly to any referenced poster, and from each poster they can access an enlarged view.

These tools make it possible to work with the collection in front of a group, in a classroom, seminar, or museum setting, as well as browsing each item individually.

At Carte, AI works as an enrichment layer on top of the existing cataloguing. Using Google Gemini models, it visually analyses each poster, generates descriptions in Spanish, identifies iconography, aesthetics, themes, and possible historical contexts, and produces semantic tags that provide new ways to search the collection.

For semantic search, each poster is converted into an embedding: a list of numbers that represents what the image shows and what its record says, so that similar posters are positioned close to each other. These embeddings are stored in PostgreSQL using pgvector and make it possible to retrieve related items even when they do not share the same authorship or descriptors.

The platform also includes a conversational assistant for the catalogue, specialised in poster art and visual culture, which suggests selections based on an objective expressed in natural language. It works on the same principle as a corporate RAG: it answers based on the collection, rather than on what the model knows on its own. We also explain this in how we built an AI chatbot for our documents.

Everything produced by the model is recorded as a suggestion and does not alter the original record. Once a professional reviews and approves it, it is stored as validated data, with a record of that review. Traceability is built directly into the data model.

The toolkits are the part of the platform that turns the collection into practical teaching material. Instead of browsing the archive item by item, the user starts with a collection and defines the parameters of the activity: educational level or group profile, duration, teaching approach, learning objectives, and any other instructions they consider relevant.

Based on this information, the platform suggests a selection of posters organised into sections, explaining the role of each work within the session. The user can request changes, replace items, and approve a final version.

Once approved, that selection becomes the single source for the toolkit: the same posters are used in the teacher's guide, presentation mode, comparison tool, downloadable PDF, and exported materials. This was an important requirement of the project, because tools that generate each document separately can end up producing guides that refer to items that are missing from the presentation.

The teacher's guide is generated using a fixed structure: session title, summary, learning objectives, preparation, step-by-step activities, questions for observation and discussion, closing and assessment activities, and possible adaptations. Each section can be rewritten using natural language instructions, and the platform keeps previous versions so they can be restored. The final document can be downloaded as a PDF containing the selected posters and their basic information.

This allows the same collection to be adapted for a primary school group, a university seminar, or a guided tour simply by changing the activity parameters.

Opening up a heritage collection means deciding which version of each image every user can access. Carte generates several variants for each poster and serves the appropriate one depending on the context and permissions: thumbnails optimised for the catalogue, protected higher-resolution versions, and watermarked images for viewing.

The original files remain in private storage, with no public URL. All access goes through controlled routes and expiring signed URLs, and delivery is adapted to each user's permissions and plan. [If confirmed:] The entire platform, including images and the database, runs in Google Cloud's European region, so the data does not leave the European Union.

With this system, the collection can be browsed, presented, and exported as educational materials without publicly exposing the original high-resolution files, allowing an institution to open up its archive while retaining control over it.

None of the above depends on the collection being made up of posters. The three data layers, semantic search, the assistant, and protected image delivery can be applied to any large catalogue where today you need to already know what you are looking for in order to find it.

An online store with thousands of poorly tagged products. AI suggests attributes based on the image and description, the team validates them in batches, and customers can search based on what they see (“straight-cut green jacket”) without knowing the product reference.

An image library. Automatic image tagging with human review, search by visual content, and delivery of each file at the resolution and with the watermark allowed by its licence.

Internal documentation. Manuals, technical data sheets, or procedures with semantic search and an assistant that answers by citing the original document, with each piece of information marked according to its source.

Each part of the stack was chosen for a specific reason related to the project:

Carte has made a collection of more than 7,500 items available to anyone who wants to work with it: it can be explored through the catalogue, organised into toolkits, and exported as materials that a teacher or mediation team can print and present on the same day.

For those responsible for the collection, the project has expanded the ways in which the archive can be accessed without giving up control over it: professional metadata remains the reference, automatically generated enrichment is clearly identified as such, and the original files remain in private storage.

We continue to work with the team on the evolution of the platform: refining semantic search based on real-world usage, expanding the capabilities of the toolkits, and preparing for the addition of new collections, as the architecture is not tied to a single archive.

Do you have a catalogue or archive that no one can find their way through? This is how we approach custom AI development: with your data, under your control, and with a person making the decisions.