Case Studies

Eoin, you have been fundamental to the development of Diotima. We knew that governance would be a cornerstone. What we didn't know was how to go about it with rigour.

Background

Diotima set out to build an AI-powered formative assessment platform tailored to the Irish secondary school curriculum, specifically targeting Leaving Certificate alignment. The challenge was significant on two fronts simultaneously.

The first was pedagogical: Ireland's curriculum is highly specific, and generic AI tools consistently fail on Irish historical context, Irish language content, and the particular assessment rubrics used by the State Examinations Commission. Off-the-shelf solutions produce plausible-sounding content that does not survive contact with an Irish classroom.

The second was regulatory. Under the EU AI Act, any AI system that analyses student responses and structures performance interpretation is classified as high-risk under Annex III, Section 3(a). That classification was not incidental to the product. It shaped every architectural and governance decision from the outset. From day one, Diotima was designed to meet those obligations, not to retrofit compliance onto an existing product.

The Formative Assessment Lifecycle

Stages 1-4 flow left to right, then the workflow wraps to stages 5-7 below. Approve and Review are teacher gates: no AI output reaches a student without passing through both. The AI proposes; the teacher decides.

What We Did

• Compliance-Grade Architecture and Curriculum Grounding

  The most consequential early decision was to treat the EU AI Act's Annex III classification as a design brief rather than a documentation task. That meant scoping the high-risk components narrowly (the Question, Answer and Rubric Generation Engine, and the AI Inference and Feedback Engine) and separating them cleanly from supporting infrastructure. Within that perimeter, every AI-generated question, answer, and rubric is explicitly grounded in curriculum topics, learning outcomes, Bloom's Taxonomy cognitive levels, and approved, licensed source materials. The architecture and the content discipline are the same idea: keep the model from drifting, give teachers the visibility to approve confidently, and give regulators a clear basis for understanding why any given item exists.

• Human Oversight as a Structural Feature, Not a Policy Statement

  Operationalising Article 14 (Human Oversight) in a way that was real rather than nominal meant enforcing oversight through the workflow itself. Teachers must approve every generated assessment item before any student sees it. AI rubric placements are always provisional. Teachers can accept, edit, override, or replace any output at any stage, and only teacher-confirmed results are stored as authoritative. Students never receive AI-generated feedback that has not been reviewed by a teacher. That is a technical constraint embedded in the system, not an aspiration in a policy document. Explainability is calibrated the same way: teachers see the full reasoning and source materials behind every output; students see performance bands progressively, after a genuine attempt, in a way that supports learning rather than gaming.

• Model Governance and Continuous Monitoring

  Model choice was treated as a governance decision, not a technical one. Candidates were evaluated against a structured benchmark suite covering knowledge and reasoning, factuality, instruction following, bias and fairness, and toxicity and safety. No student data is used for model training at any stage. Model versions are logged per inference to support audit and incident investigation. The teacher approval and rejection workflow doubles as a continuous monitoring mechanism: every rejection is a signal about model performance in a real Irish classroom, fed back into the risk register and model governance process. Compliance, in this design, is something the system generates evidence for as a byproduct of normal operation.

• Content Licensing Strategy and Investor Narrative

  As the platform matured, content licensing with Irish publishers became a key risk. We identified this early and advised on how to structure publisher agreements explicitly covering storage, processing, and derivative use of licensed materials within the scope of Diotima's purpose. This became part of the investor narrative: the platform's differentiation (Irish-specific, curriculum-aligned, teacher-validated, and compliance-grade) was articulated not just as a product thesis but as a regulatory moat. In regulated education markets, compliance-by-design gives buyers access to institutions that non-compliant products simply cannot legally serve.

Background

A top-five global custodian bank managing over $45 trillion in assets under custody needed to move beyond ad-hoc data science experiments and build a mature, governed AI capability. The challenge was not technical ambition. It was institutional: how to deliver commercially impactful AI solutions across a highly regulated financial institution while maintaining the trust of clients, regulators, and internal stakeholders.

The mandate was to build and lead a global team of data scientists and engineers, establish governance structures for responsible AI development, and deliver measurable business outcomes across all lines of business, from custody settlement and asset servicing to client analytics and operational efficiency.

What We Did

• Built and Led a Global AI Team

  Recruited, developed, and managed a worldwide team of data scientists and engineers across multiple geographies. Established a collaborative operating model spanning business units, technology, risk, and compliance, ensuring AI solutions were commercially aligned and institutionally supported from inception.

• Implemented AI Governance Before It Was Required

  Designed and implemented a rigorous governance model for AI development and deployment, covering model risk, bias testing, explainability, and ethical review. This was established years before the EU AI Act made such governance mandatory, giving the institution a structural advantage when regulatory requirements emerged.

• Delivered 30+ Enterprise AI Solutions

  Collaborated with all lines of business to identify and implement transformative AI solutions spanning predictive analytics, anomaly detection, automation, and collective intelligence. This included Custody Predictive Trade (a first in the custody settlement industry), alongside solutions in client analytics, regulatory reporting, and operational risk.

• Established a Digital R&D Hub

  Secured an €8 million investment to establish a Digital R&D Hub in Dublin, recognised by IDA Ireland. The hub became a centre for innovation in AI, data science, and emerging technologies, including quantum computing, and positioned the institution as a leader in financial services AI innovation.