Location: Dublin (NOCC)
Salary Range
€89,180.00 - €165,620.00
Job Description Summary
Location: Dublin, Ireland - Hybrid
Relocation is not offered for this role. Please apply if this location is accessible to you.
This is a hands-on senior role in the Data Science & AI (DSAI) team within Process Innovation, part of Strategy, Program & Portfolio Operations (SPPO) in Novartis Global Drug Development. Your focus is how Development works: the programme management, portfolio, governance and planning processes that move medicines from decision to decision. You will build, evaluate and scale AI skills for the assistants colleagues already use every day, such as Microsoft Copilot Cowork and Claude Cowork. These skills are built in language rather than code: instructions, checklists, templates and knowledge that tell the assistant how to do a real piece of SPPO work well.
The role centres on three questions that decide whether these skills create value. What knowledge do they need, and how do we organise it so every skill draws on the same trusted sources? How do we review and evaluate skills at scale, rather than one by one? And how do we earn enough trust to let skills act (drafting and updating documents, records and messages) rather than only advise? We are looking for an engineer who wants to code less and solve more business problems. You will build with AI tools such as Claude Code, GitHub Copilot or AWS Kiro, and your engineering foundations are what let you design the system flow and the knowledge layer underneath each skill.
Major accountabilities
Skill design: building in language
Design and build skills for enterprise AI assistants (Microsoft Copilot Cowork, Claude Cowork) in which language is the programming interface: instructions, function-specific checklists, worked examples, templates and references to trusted sources.
Translate SPPO processes and decisions, such as tollgate and governance reviews, development plan checks, portfolio reporting and programme documentation, into precise, testable skill specifications, working directly with the programme leaders and process owners who run them.
Design how work flows across skills, agents and tools: what each step sees, what it produces, and where a human checks or decides. Apply agent-to-agent (A2A) patterns where skills in Copilot, Claude and other agents need to hand work to each other, with clear contracts for what is passed and who is accountable.
Use AI building tools (Claude Code, GitHub Copilot, AWS Kiro) to author, test and maintain skills and the code around them, such as connectors, data preparation and evaluation harnesses, keeping each skill versioned, documented and correctly classified for the data it handles.
Recognise when a need has outgrown a skill, and hand it over to the agentic engineering, product or platform teams with a clear evaluation record.
Knowledge that fuels the skills
Organise the knowledge skills draw on into a shared, trusted layer: checklists, definitions, templates, guidance and reference documents, curated once and reused by every skill rather than copied into each.
Decide what goes on the platform: specify which trusted sources skills need and how they are exposed, through Model Context Protocol (MCP) servers and approved connectors that DDIT builds and runs. Design the data structures and patterns behind this layer so skills reach the right sources in SharePoint, Teams and business systems with the right access, provenance and freshness, and so each source serves many skills rather than one use case.
Keep this knowledge maintained, owned and versioned, so that a change to a checklist or guideline flows to every skill that uses it.
Review and evaluation at scale
Build evaluation for every skill: test cases drawn from real documents, expert reference answers, and structured review (including AI-assisted grading checked against human experts), re-run whenever a skill changes.
Measure what matters for the decision: correctness against expert judgement, coverage, hallucination and omission rates, consistency, and honest reporting of uncertainty.
Make evaluation repeatable, so that review scales with the number of skills rather than the number of reviewers, and monitor skills in use for quality and drift.
Contribute to DSAI evaluation standards, working with the Director, AI Systems Engineering & Evaluation.
Trust for skills that act
Define the path from advising to acting for each skill (read, draft, propose, act with approval, act on its own) and the evidence needed to move up a level.
Set the guardrails for skills that write back, such as updating a document, a SharePoint list or a tracker, or sending a message: least-privilege access, previews, human approval points, reversibility and a clear audit trail.
Ensure systems that act meet GxP, data privacy and Responsible AI requirements, working with Quality, Compliance and DDIT.
Forward deployment and the build factory
Embed with SPPO teams (Global Program Management, portfolio, governance and operations) and the Process Innovation team, starting from the process and the decision rather than the technology: use design thinking and Lean methods to understand the work, find where skills can change how it runs, then land them in real use and measure their adoption and impact.
Coach early-career scientists on rotation in AI-native building and evaluation, and share practice internally and externally through publications, conferences and communities of practice.
Run skill building as a factory rather than one-off support: shared intake and triage, reusable components and patterns, and a maintained register of every system with its business owner, evaluation status and measured value.
Job description provided by Novartis.
View the original posting