AI-MUSE — Venice · Larnaca · İstanbul — pilot applications open
Your collection.
Every language.
We take what your institution already holds — images, video, documents, digital text — read it, have it checked by people who know archival material, then translate and voice it. In Europe, the AI included. Free for twenty-four months, and the rights to your material never move.
The technology is demonstrated on real archival material — documented below, with the numbers. The full service goes live with the pilot programme: applications are open. Step 2 application due by 31 August 2026.
Why this exists
The material is catalogued. It is the reading and the publishing that never happen.
An archive funded by a single municipality has a catalogue, boxes described decades ago, and rare books nobody has ever scanned. Between that and a wall label, a catalogue record or an audio guide in four languages sits a cost per page no one has ever been able to justify — so the fund stays closed and the same twenty objects get shown.
Which is why a pilot has to end with something you can put in front of your board, and not with a demonstration.
What we do
A collection goes in. Something publishable comes back.
Recognition is the part of this that is already solved, and it was solved in Europe: we adopt those tools instead of competing with them. What is not solved is everything around them — material arriving in six different formats, text that needs an expert eye before anyone relies on it, and then the same content in as many languages as your visitors speak. That chain is still done by hand everywhere, and it is the chain we automate.
Six stages, one of which is a person from your institution saying yes.
What already exists
An entire year of a newspaper, read by machine and mapped.
A complete run, not a sample. Giustizia e Libertà, the Venetian weekly of the Partito d'Azione: 34 consecutive issues, 70 pages, from 3 June 1945 to 20 January 1946, directed by Agostino Zanon Dal Bo and printed under Allied P.W.B. licence. Six columns a page, segmented before reading. 292,771 words in 4 hours 23 minutes of compute, on one machine in Venice, with no cloud bill.
Then the map. 951 mentions in the text linked to 506 distinct entities — people, places, organisations. Below: every page, every issue and every entity as one graph. The dense knots are what the year was about: the Partito d'Azione, Venice, Rome, the Comitato di Liberazione Nazionale, Silvio Trentin.
609 nodes, 1,866 links — click to open full size.
What matters is not that it worked. It is that the output says what it is: suggested corrections are flagged and not applied; blocks too degraded to read are marked as unreadable instead of being filled in with plausible Italian; and every page carries a provenance block — source scan, recognition model, date, columns, compute time — so any line can be taken back to the page it came from.
And the entity links are not finished work: we checked all 506 against Wikipedia and Wikidata, and 26% need a curator — surname-only references that land on disambiguation pages, and identifications that are confidently wrong. That figure is the argument, not the embarrassment: it is the size of the human step, measured instead of asserted.
And a dictionary, for scale. 7,444 pages of a nineteenth-century dictionary and 126,660 structured entries, in about 140 machine hours on a single 8 GB consumer graphics card, with no cloud bill. Residual transcription error 0.5–2%, deliberately left in and published, so every entry can still be checked against its source page.
The dictionary figures are as declared by the firm that ran the campaign — our own IT lead's — and we have not independently re-measured them. A result whose margin of error is unknown is not a result, and that goes for ours too.
The human version of this newspaper already exists, and it is better than ours at what it does. IVESER catalogued the title in Un secolo di carta, its census of Venetian periodicals — format, director, print run, which four libraries hold it. And the clandestine sheets of 1943–45, which are a different body of material from this legal post-Liberation weekly, are in the exhibition «Veneziani!!!» Fogli resistenti. Our corpus does not replace either: it adds the searchable text and the links, and it agrees with the catalogue on every point we could check.
The full story of that campaign — the scans, the supercomputer and the entity graph — is told in L’Ingegneria Civile della Memoria on the ELAN42 blog (in Italian).
The workflow
- 01
- CollectImages, video, documents — and the digital text you already hold.
- 02
- ReadHistorical print and degraded print: demonstrated. Non-Latin scripts and handwriting: models trained on your own material, and the target of the pilot rather than a capability we have today.
- 03
- CheckA critical eye trained on cultural and documentary material. Nothing moves on unread.
- 04
- Your curator decidesNothing carries your institution's name without their approval.
- 05
- Translate and voiceText and audio, in any language your visitors speak.
- 06
- Hand overThe files are yours, in open formats no supplier controls: IIIF, METS, ALTO, PAGE XML, TEI, Dublin Core, EDM.
Taking part
What you sign, what stays yours, and what this still cannot do.
What you sign is one page. A letter of support saying your institution thinks the work is worth doing and is willing to test it. It is not binding: it commits you to no purchase, no exclusivity, and no obligation to continue when the pilot ends. If the project is not funded, the letter expires with it.
What it costs during the pilot is nothing. Institutions that join now are free for twenty-four months — the pilot year and the year after it. Afterwards AI-MUSE becomes a commercial product on a monthly licence — a platform fee, with one-off charges for very large digitisation batches and small batches inside the base plan. That price does not exist yet, and we would rather write this sentence than invent a figure your board would hold us to. Whatever it turns out to be, taking part in the pilot never obliges you to pay it.
What we need from you. A set of items you choose — tens of pages, not thousands, and coherent enough in hand or typeface that a model can be trained on them. Images: a backlit scan is ideal, but an overhead scanner, or even a careful phone capture, is enough as long as the letters are legible. A curator or archivist who reviews the output and says yes or no. And honest feedback, including the kind we will not enjoy — your corrections are what trains the model, which is the part of this that is genuinely mutual.
What stays with your institution. The rights to your material, entirely: we acquire none, and we do not want any. The rights status of every item is recorded when it enters the pipeline, not when something gets published. The output files, in open formats no supplier controls — IIIF, METS, ALTO, PAGE XML, TEI, Dublin Core, EDM — in your own systems, yours to keep. The recognition models trained on your hands and your scripts, which stay with you and keep working after the project ends. And the measured result: accuracy per holding, published with its error margin, so what you show your board is a number and not an impression.
And a finished thing, not a demonstration. A pilot inside a European project should leave the institution with a small completed project of its own — a fund, a volume, a set of manuscripts, done end to end and publishable under your name. That is the commitment, and it is the one to hold us to.
If you would rather not run anything yourself, we can operate it as a service instead, streaming the audio and the multilingual text to your site, your app or your galleries, on European hosting. You can start with one and move to the other; nothing is locked. And where material cannot lawfully leave the building at all — depositor agreements, licensing, confidentiality — it does not have to: the installation comes to you.
What it cannot do yet
Each of these is a real limit, on the record. Three of them are what the pilot exists to attack, which is not the same as having solved them.
- —
- Manuscript at scaleDemonstrated on historical print, including degraded cyclostyled print on hand-cut stencils. Handwriting is the pilot's target, not a capability we hold today: a limited, stylistically coherent set of manuscripts, trained per hand. An archive that is 95% manuscript should read that sentence twice before signing.
- —
- Faded eighteenth-century ink, and letterpress copiesWhere the original is barely legible to a person, it is not legible to a model either. We measure it on a sample and give you the number before you commit the fund.
- —
- Non-Latin scripts across a whole holdingGreek, Cyrillic, Arabic, Kufic: handled in principle, not yet proven at the scale of a full collection. Same rule — a sample, a measurement, a number, before you decide.
- —
- Strong dialect in recorded speechAutomatic transcription of heavy dialect does not work, and we are not going to pretend otherwise. Standard-language audio is fine; Cypriot dialect, or Sardinian, is not. Oral sources are a research interest of ours, not something we will put in a deliverable.
- —
- A whole fund in twelve monthsThe pilot is scoped in items, not in collections. Twenty thousand volumes is a programme, not a pilot. The honest page ceiling per institution is a figure we are still calculating, and you will get it before you are asked to choose the material.
Apply as a pilot institution
Tell us what the material is and what state it is in. If the answer is that we cannot read it, that is the answer you will get — with the measurement that says so.
The button opens your own e-mail client with the application ready to send to info@ai-muse.eu — the form itself stores nothing on this site. The site uses Google Analytics for anonymous visit statistics. If you prefer, write to that address directly.
European end to end
The AI included — and we did not have to compromise to get there.
The best handwriting recognition in the world is already European and already open. It is built by READ-COOP, an Austrian cooperative with more than 250 private and institutional co-owners; its engine, PyLaia, is open source under an MIT licence and is maintained by Teklia, in France. We do not try to beat ten years of that work. We adopt it through its own API, and we build the part that is missing — which is why an institution that has already invested in Transkribus loses nothing by working with us.
Processing, storage and the models themselves stay in Europe, on open-source models and European compute. Nothing is sent to a third country to be read.
That compute is not a figure of speech. AI-MUSE has been granted compute resources through the IT4LIA AI Factory (EuroHPC), project EHPC-AIF-2026PG01-1068. From 24 August 2026 the heavy recognition batches run on Leonardo Booster at CINECA, Bologna — the Italian public supercomputer, funded by EuroHPC and the Ministry of University and Research, twelfth in the TOP500 list of June 2026.
And the bottleneck here is procedure, not compute. Before the allocation this ran on one 8 GB graphics card in our office in Venice, which is what makes the cost per page predictable. What the allocation buys is the room to attempt handwriting at scale — not, by itself, the ability to read it.
Your dashboard comes with a chatbot for working with your own material, and it runs on the same European stack — not on an API somewhere else.
And where material cannot lawfully leave the building at all — depositor agreements, licensing, confidentiality — it does not have to. The models trained on your collection stay with you, and keep working after the project ends.
Who is already in
Three countries, seven signed letters, and the roles written out.
Every institution below chose its own material and keeps the rights to it. The terms are the ones set out above — the same for the next one to join.
Venice, Italy · Lead Partner
The workflow, the models, the European infrastructure.
elan42.com
Larnaca, Cyprus · Project Partner
Multilingual validation, the Cypriot pilot, dissemination.
elitetutors.org
CREAapsItaly · Associated Partner
Associated Partner and research organisation. The metadata scheme, the taxonomy, and the expert check on meaning, context, bias and omission. Letter of support signed, 26 August 2026.
creasiena.it
Venice, Italy · Associated Partner, hosts the Italian pilot
Two holdings — the clandestine Resistance press of 1943–45 and the Agostino Zanon Dal Bo papers, nine boxes catalogued and never digitised — and the curator who approves what is published. Letter of support signed.
iveser.it
Larnaca, Cyprus · hosts the Cypriot pilot, with the Pierides Foundation
Material from the Pierides Foundation and family archives, late 19th and 20th century, printed and handwritten, with the Museum's own curatorial authority approving what is published. Letter of support, 12 August 2026; pilot corpus committed, 23 August 2026.
bankofcyprus.com
Larnaca, Cyprus · supporting institution
A family archive of more than 5,000 letters from the 1890s, in Greek and Ottoman-era Arabic, still in boxes after fifteen years. Letter of support, 20 August 2026.
Türkiye · Associated Partner, commercial channel
Brings the workflow to Turkish museums and archives under a signed commercial agreement — and brought in the third pilot, NOAH. Letter of support, 25 August 2026.
İstanbul, Türkiye · Associated Partner, hosts the Turkish pilot
Handwritten material of the institution's own choosing, with NOAH reviewing every output before anything is published under its name. Letters signed, 26–27 August 2026.
Venice, Italy · investor interest — not a consortium member
Venice's deep-tech venture builder, born after the 2019 Aqua Granda flood, signed a letter of interest on 27 August 2026: evidence that the service has a market, not only a grant.
venisia.com
Behind it
AI-MUSE is a startup by ELAN42 — E42 S.r.l., a digital agency in Venice working on AI, cultural heritage and European infrastructure.
Twelve publicly funded projects delivered since 2018, and more than ten years in the trade. The multilingual layer is not a promise either: a version of it already runs in production for our hosting clients, on our own GPU server, with no content leaving our infrastructure and no third-party API.
Two of the things this project promises to measure, we are qualified to measure. AI-MUSE commits to reporting the social return and the carbon cost of the pilot: its CEO holds a CPD-certified SROI Practitioner qualification, and the company has been carbon neutral since 2017 through LifeGate's Impatto Zero — reforestation in Madagascar, 8,778 m² of woodland financed in the Parco del Ticino, protected Amazon forest through Foreste in Piedi. Electricity is 100% renewable, the servers run in hydroelectric data centres, and the hosting is verified by the Green Web Foundation. The full record is public. Numbers nobody can check are worth nothing.
One of 16 proposals out of 302 selected for Step 2 of the CR.EU.IN. HERITAGE Cultural Heritage Innovation Grants; ten will be funded. The Step 2 application closes on 31 August 2026 and the project is not yet funded. We say this here because a letter of support is worth knowing the odds on: if the project is not selected, your letter expires with it and costs you nothing.
Partners and networks
Who is in the project — and where the Lead Partner already sits.
Elite Tutors — Project Partner
CREAaps — Associated Partner: research organisation · ENOVA — Associated Partner: the Turkish channel.
And these are not project partners: VeniSIA — the investor interest — and ELAN42's own memberships. They are how the results travel once there are results.
ELAN42 is a member of Venetian Cluster, the Regional Innovation Network (RIR) for cultural heritage and the creative industries of the Veneto Region.