D3.1. Design of the AI Gateway

Executive Summary Context. The Serbian AI Factory Antenna (SAIFA) is a HORIZON-JU-EuroHPC project (Grant Agreement No. 101263280) that establishes a national entry point to artificial intelligence and high-performance computing for research and clinical institutions, industry, the public sector and individual users. The AI Gateway is developed in Work Package 3, Task 3.4, as the unified access point for AI resources, tools and services. Deliverable D3.1, due in M03, provides the comprehensive initial plan for its development. Scope. The design covers access to computing environments, datasets, pre-trained models, applications, reusable tools, modular workflows, training materials and support services. It also addresses authentication, resource scheduling, monitoring, secure sharing and interoperability with national infrastructure, Pharos, IT4LIA, the AI-on-Demand Platform and other AI Factory services. The design sets out a target direction and an evaluation process. It does not predetermine the final implementation scope, architecture or division of responsibilities, nor does it commit the project consortium to delivering every described capability or workflow in the beta or subsequent releases. Design approach. The Gateway is conceived as a modular integration and control layer rather than a monolithic platform. Mature components are reused for identity management, cataloguing, notebooks, model management, execution, storage, monitoring and learning. SAIFA-specific development is concentrated on the unified portal, common metadata and authorisation rules, quota and accounting logic, the HPC bridge, audit, and federation adapters. Users and use cases. The functional scope is illustrated through 21 Domain Use Stories covering language and culture, biomedicine and health, energy and manufacturing, environment, and cross-domain adoption. A representative set of system use cases specifies the principal interactions for authentication, discovery, data and model management, compute jobs, inference, experiments, training, federation and support. These descriptions remain subject to validation and prioritisation with domain owners, infrastructure operators and user-support teams. Technology and risk. Candidate technologies are evaluated through short proofs of concept, mandatory elimination criteria and comparative assessment of usability, integration effort, operational complexity, performance, maturity, compliance and sustainability. The accompanying risk register combines proposal-level risks with Gateway-specific security, integration, data-integrity and operational risks, and links them to preventive controls and acceptance evidence. Roadmap. The contractual path is D3.1 in M03, access to national infrastructure in M03, access and use policies and the online beta version in M12, and D3.4 - AI gateway functionalities and features - in M18. Task 3.4 continues through M36, covering further integration, hardening, interoperability, validation and operational improvement. Development and preliminary testing are carried out on the PARADOX and ITE infrastructures, followed by validation and benchmarking using the Greek and Italian AI Factory infrastructures. Outcome. Together, these documents form a coherent design baseline that links the project structure, domain needs, system behaviour, candidate technologies, delivery plan and risk controls. The baseline supports implementation while preserving the flexibility needed to incorporate validated user requirements and evidence from the planned proofs of concept.