PASTFORWARD.AI

The project

Moving forward while keeping the past alive using AI

PASTFORWARD.AI is a 48-month research and valorisation project of KIK-IRPA, the Royal Institute for Cultural Heritage, funded by the Belgian Science Policy Office (BELSPO) under the Policy 4 Science programme, call 2024–2025.

In a nutshell

Existing AI, applied with care

Developing new AI from scratch is expensive, demands rare skills, and raises data privacy concerns. Relying on large commercial platforms brings its own problems. PASTFORWARD.AI takes a third route: it identifies, evaluates and applies readily available open-source AI models to the needs of a cultural heritage institute.

The goal is to strengthen the work of KIK-IRPA's three core departments: better research quality, more efficient operations, and more accessible collections, particularly through BALaT, the institute's online portal to over a million photographs and library records. Along the way, the project improves the FAIRness of our data: findable, accessible, interoperable and reusable.

The project focuses on inference, meaning it uses pre-trained models rather than training new ones. Fine-tuning is only considered where a proof of concept genuinely needs it. Everything runs on the institute's own hardware, because some of our datasets are confidential and cannot leave the building.

Objectives

What we want to achieve

The project formally tracks 26 research objectives. They boil down to six commitments.

01

Run the project well

Solid governance with a stakeholder committee of external experts, quality reviews on every major deliverable, a living data management plan and proper data security.

02

Do it responsibly

Develop and apply a framework for assessing AI-specific ethical issues such as bias, copyright and the risk of misrepresenting heritage and the communities behind it.

03

Know the landscape

Build a curated list of open-source AI models suited to heritage tasks, and learn from international initiatives so we do not repeat mistakes others already made.

04

Choose wisely

Collect real use cases from colleagues and partner federal scientific institutions, prioritise them together, and map the best models to the best use cases.

05

Prove it works

Build working proofs of concept on our own data and infrastructure, test them against clear success criteria, and document what works and what does not.

06

Plan ahead and share

Turn the results into a strategic AI roadmap for the institute, with concrete next steps, and share the knowledge, code and lessons with the whole sector.

Methodology

Foundations, experimentation, reflection

The methodology unfolds in three interlinked phases. Each phase builds on the outputs of the previous one, forming a path from early exploration to prototyping, evaluation and long-term strategy.

Diagram of the three project phases: Foundations (WP1 to WP3), Experimentation (WP4 and WP5) and Reflection (WP6 and WP7), with icons for project governance, AI model experimentation and strategic planning.
fig. 1 — The structured progression of the PASTFORWARD.AI methodology.
1WP1–3 · M1–M15

Foundations

Governance, data management and ethics are set up first, alongside a focused survey of open-source AI models and of international AI initiatives in cultural heritage. This phase produces the evidence base for everything that follows.

2WP4–5 · M6–M42

Experimentation

Models and use cases are selected and matched through workshops, interviews and multi-criteria analysis. The strongest pairings become proofs of concept, developed and iteratively tested on the institute's own GPU infrastructure.

3WP6–7 · M40–M48

Reflection

The proofs of concept are evaluated, lessons are synthesised, and the institute's readiness for AI is assessed honestly. The result is a strategic AI roadmap and a set of concrete recommendations, shared widely.

Timeline

48 months at a glance

The two coordination work packages run for the whole project, and dissemination starts almost immediately. The research work packages hand over to each other as the phases progress.

Work packages

Seven work packages, 42 deliverables

Each work package below can be unfolded to show its tasks and deliverables, with the project month in which each deliverable is due.

WP1 Coordination, project management and reporting M1–M48

Keeps the project on course: planning, risk management, reporting to BELSPO, and quality reviews of every major deliverable. A stakeholder committee of external experts meets at least four times during the project to challenge our direction and keep the work relevant.

  • T1.1Project governanceM1–M48
  • T1.2Project management and quality assuranceM1–M48

Deliverables

D1.2.1Initial reportM3
D1.1.1Minutes of the stakeholder committee kick-off meetingM10
D1.2.2First annual activity reportM12
D1.1.2Minutes of the second stakeholder committee meetingM22
D1.2.3Second annual activity reportM24
D1.1.3Minutes of the third stakeholder committee meetingM34
D1.2.4Third annual activity reportM36
D1.1.4Minutes of the final stakeholder committee meetingM46
D1.2.5Final reportM48
WP2 Research data management and ethics M1–M48

Builds the data and ethics backbone of the project: a living data management plan aligned with the FAIR principles, a backup and security strategy following the 3-2-1 rule, and an AI ethics assessment framework that is applied and refined throughout the project. An ethics issue log keeps a transparent record of dilemmas and how they were resolved.

  • T2.1Development of a data management plan and compliance with FAIR principlesM1–M48
  • T2.2Implementation of data security measures through backup policiesM1–M48
  • T2.3Upholding responsible research practices and addressing AI-specific ethical principlesM3–M48

Deliverables

D2.1.1Data management planM6
D2.3.1Initial AI ethics assessment framework and responsible research guidelinesM12
D2.2.1Report on data security during the projectM48
D2.3.2Final report on responsible research practices, AI ethics framework and logM48
WP3 Focused survey of AI models and initiatives in cultural heritage M1–M15

Maps the landscape before anything gets built. The team shortlists promising open-source models for the five application areas, using strict criteria such as licensing, ease of deployment and minimal fine-tuning needs. In parallel it studies international AI initiatives in cultural heritage, with the AI4LAM community as a key source, and distils everything into actionable lessons.

  • T3.1Define the research scope and prioritisation methodologyM1–M4
  • T3.2Identify and document a shortlist of promising AI modelsM2–M9
  • T3.3Identify and document key international initiatives and learningsM4–M12
  • T3.4Analyse findings and synthesise actionable lessonsM9–M15

Deliverables

D3.1.1Research scope, prioritisation criteria and methodologyM4
D3.2.1Report on shortlisted and prioritised AI modelsM9
D3.3.1Report on international AI initiatives and best practicesM12
D3.4.1Actionable lessons and recommendations for the proofs of conceptM15
WP4 Selection, prioritisation and mapping of AI models and use cases M6–M22

Turns the broad survey into concrete targets. Colleagues across departments and partner federal scientific institutions bring in real use cases through workshops, interviews and surveys. Models and use cases are then scored, prioritised and mapped to each other, ending in a shortlist of model and use case pairings recommended for proof-of-concept development.

  • T4.1Selection and prioritisation of AI modelsM9–M16
  • T4.2Identification and documentation of potential use cases with other federal scientific institutionsM6–M15
  • T4.3Evaluation and prioritisation of potential use casesM15–M18
  • T4.4Mapping of prioritised models to prioritised use casesM17–M20
  • T4.5Refinement of recommendations for the proofs of conceptM20–M22

Deliverables

D4.2.1Draft inventory of potential AI use casesM15
D4.1.1Report on selected and prioritised AI modelsM16
D4.3.1Inventory and prioritisation of potential AI use casesM18
D4.4.1Model-to-use-case mapping matrix and fit assessmentM20
D4.5.1Recommendations for proof-of-concept developmentM22
WP5 Development of proofs of concept M22–M42

The hands-on phase. Each selected pairing gets a detailed plan with measurable success criteria, curated datasets, and a reproducible technical environment on the institute's own servers. Development is iterative, with testing and refinement running in parallel, and everything is documented so the results can be reproduced and reused. The finished proofs of concept are demonstrated to stakeholders.

  • T5.1Detailed planning and design for each selected proof of conceptM22–M24
  • T5.2Data preparation and curationM23–M27
  • T5.3Technical environment setup and model integrationM24–M28
  • T5.4Development and implementation of the prototypesM26–M38
  • T5.5Testing, evaluation and refinementM28–M40
  • T5.6Documentation of the development process, challenges and resultsM26–M41
  • T5.7Demonstration and presentation of the proofs of conceptM39–M42

Deliverables

D5.1.1Detailed proof-of-concept plansM24
D5.2.1Curated datasets for developmentM27
D5.3.1Development and testing environments establishedM28
D5.4.1Demonstrable proof-of-concept implementationsM38
D5.5.1Test results and refinement logsM40
D5.7.2Recommendations for the WP6 evaluationM40
D5.6.1Technical documentation and evaluation report per proof of conceptM41
D5.7.1Demonstration materials and session recordsM42
WP6 Evaluation of the proofs of concept and strategic roadmap M40–M48

Looks back to plan forward. The proofs of concept are evaluated against their success criteria and their potential to scale. Lessons from the whole project are synthesised, the institute's readiness for AI is candidly assessed, and everything feeds into a strategic AI roadmap with concrete recommendations for the next steps.

  • T6.1Comprehensive evaluation of the developed proofs of conceptM40–M43
  • T6.2Synthesis of project findings and lessons learnedM41–M44
  • T6.3Assessment of long-term AI integration potential and organisational readinessM42–M45
  • T6.4Development and prioritisation of the strategic AI roadmapM44–M47
  • T6.5Formulation of recommendations for next stepsM46–M47
  • T6.6Reporting and dissemination of the evaluation and roadmapM47–M48

Deliverables

D6.1.1Aggregated evaluation data and stakeholder feedbackM43
D6.2.1Comprehensive evaluation reportM44
D6.3.1Organisational AI readiness assessment and gap analysisM45
D6.4.1Strategic AI roadmapM47
D6.5.1Recommendations for immediate next stepsM47
D6.6.1Final project report and recommendationsM48
WP7 Dissemination and valorisation M2–M48

Makes sure the results leave the building. This website and the blog are part of it, alongside tailored materials for heritage professionals, researchers and policymakers, presentations at conferences and workshops, peer-reviewed publications, and the open-source release of the proof-of-concept code. Networks such as AI4LAM, ICOM and the Europeana Network Association help the results travel.

  • T7.1Valorisation activities onlineM2–M48
  • T7.2Dissemination towards target groupsM12–M48
  • T7.3Scientific valorisation of the project resultsM18–M48

Deliverables

D7.1.1Project website and online presence establishedM2
D7.1.2Publicly accessible digital project outputsM48
D7.1.3Open-source proof-of-concept code with technical documentationM48
D7.2.1Report on dissemination activities and outreachM48
D7.3.1Scientific publications submittedM48
D7.3.2List of scientific presentations at conferences and workshopsM48
A note on expectations. The proofs of concept exist to test feasibility and to learn, not to ship production systems. They show what is possible, what it costs, and what it would take to scale, which is exactly the evidence the strategic roadmap needs.