S2GAIIA: IA Generativa Especializada e Sustentável para Aplicações Industriais

The high-level goal of S2GAIIA is to develop a methodological framework to manage the lifecycle of specialized generative models in complex socio-technical systems, addressing representative categories of organizational situations. This high-level goal can be hierarchically broken down below. G1: Develop a methodology for streamlining the adoption of GAI by organizations, facilitating the development of smaller domain/task-specific models and applications to support organisational processes and decision-making. This goal strives for sustainability, efficiency, and access. It is about making models smaller and more specialized, making them easier to develop and use within the organizational context, making the technology more accessible. The activities contributing to this goal are Activities 2 and 4. G1 is further refined into two sub-goals: G1.1: Develop context-specific datasets, evaluation benchmarks and metrics that can accurately evaluate model quality and alignment within the organization. There is a clear value and opportunity in developing domain-specific datasets enriched with contextual information about the organisation. However, these datasets must be accompanied by appropriate and domain-specific evaluation benchmarks that allow the assessment of model performance from a socio-technical perspective, both task-oriented and values-oriented. G1.2: Automate the process of developing, fine-tuning or (re-)using specialized GAI models in specific organizational contexts and across organizations. This is fundamental for minimizing the necessary resources and skills to develop specialized GAI models and applications through the development of optimal search processes, guided by techniques such as meta-learning or optimization. G2: Develop a novel architecture that uses relevant contextual information about people, processes, and systems of the organisation to inform and support the lifecycle of GAI models and applications, especially in what concerns training, fine-tuning and (re-)using models. This goal focuses on a human-centric and organisational perspective to GAI, instead of the purely technological that has been followed hitherto. It is about modelling and using information about people, systems and organisations to develop more aligned models with a stronger link to context. This goal will be achieved through the work developed in the Activities 1 and 3. G2 is expanded into two sub-goals: G2.1: Model relevant context information about people, processes and systems in the organisation in order to support its integration and use in the lifecycle of GAI models and applications. This is about recognizing the importance of the organization’s context and crystalizing it into appropriate semantic artefacts and data structures that facilitate its use. G2.2: Develop a method, to be embedded in the relevant organisational processes, to integrate relevant contextual information about the organization into developing GAI models and applications, ensuring factual accuracy, grounded outputs, and connection to existing enterprise systems. This goal builds on G2.1 to streamline the use of contextual information about an organization in the context of GAI pipelines and explore it to develop more context-aware models. These goals are, as we understand them, ambitious enough to lead to impactful scientific contributions, and realistic enough to allow for the development of multiple PoCs in the form of prototypes for specific use cases.