A KIK-IRPA research project · funded by BELSPO
Moving forward while keeping the past alive, using AI
PASTFORWARD.AI explores how open-source artificial intelligence can help cultural heritage institutions open up their collections. We test what already works, build proofs of concept on real heritage data, and turn what we learn into a practical AI roadmap.
Why this project
A lot of heritage, not enough hands
KIK-IRPA, the Royal Institute for Cultural Heritage in Brussels, studies, documents and restores Belgium's cultural heritage. Over a century of work has produced millions of photographs, intervention files, reports and records, much of it digitised and available through platforms like BALaT. Making all of that truly findable and usable is more work than any team can do by hand. That is where AI can help, if it is applied carefully.
Open source first
We do not train new models from scratch or lock ourselves into commercial platforms. The project surveys, selects and applies existing open-source AI models that institutions can actually afford to run.
Local by design
Models run on our own hardware. Sensitive collections and confidential files never leave the institute, which keeps us in control of privacy, copyright and data sovereignty.
Ethics built in
An AI ethics assessment framework is developed early in the project and applied to every model and use case, from bias and copyright to the risk of misrepresenting heritage.
Five applications
What we are testing AI on
The project focuses on five concrete application areas, chosen because they answer real needs across the institute's departments. Each one is explored through working proofs of concept on our own collections.
Reverse image search
Find related photographs and artworks by visual similarity instead of keywords, even when only a detail matches.
Read more → 02Document Q&A with RAG
Ask natural questions across thousands of reports and files, and get answers grounded in the documents themselves.
Read more → 03Metadata enrichment
Let image recognition suggest descriptions, subjects and tags, so collections become searchable in depth. A human always validates.
Read more → 04Contextual translation
Translate catalogue records between Dutch, French and English with the full record as context, and heritage vocabularies as the reference.
Read more → 05Advanced OCR
Get reliable text out of digitised documents, including layouts, schemas and images that standard OCR engines give up on.
Read more →Built as proofs of concept
Each application is tested on real KIK-IRPA data, evaluated with the people who would use it, and documented so other institutions can learn from the results. Code is published open source.
How it unfolds
Three phases over 48 months
The project runs for four years and builds up step by step: first understanding the landscape, then experimenting hands-on, and finally turning the results into strategy.
Explore and prepare
Set up project governance, data management and the AI ethics framework. Survey the landscape of open-source AI models and learn from international initiatives in cultural heritage.
Select and build
Match the most promising models to real use cases gathered with colleagues and partner institutions, then develop and iteratively test proofs of concept on our own infrastructure.
Evaluate and share
Evaluate the proofs of concept, distil the lessons learned into a strategic AI roadmap for the institute, and share results, code and recommendations with the wider sector.
Working in the open
Everything we learn gets shared
PASTFORWARD.AI is a valorisation project: its purpose is not just to build things, but to make the knowledge usable by others. KIK-IRPA is a member of the international AI4LAM community, where institutions exchange experience on AI for libraries, archives and museums.
Blog and updates
Progress, milestones and honest lessons learned are posted on the project blog as the work happens.
Presentations on Zenodo
Slides and posters from conferences and workshops are archived on Zenodo and linked from the downloads page.
Open-source code
The proof-of-concept code and its technical documentation are released under open-source licences in public repositories.
Public deliverables
Reports, the ethics framework and evaluation summaries are published on this site as they are completed.