Thesis work - Context Engineering for AI Agents in Enterprise Analytics

As a thesis student with us, you’ll have the opportunity to put your theoretical knowledge into practice while gaining unique insight into our operations. You’ll become part of our team, contribute to our development, and acquire valuable experience that strengthens both your skills and your CV. We look forward to hearing your ideas and supporting your growth!

The sustainable transformation begins with us

We are facing our greatest challenge yet - ensuring that Sweden and Europe have what it takes to transform. So that we can live the lives we want, both today and tomorrow. In the same way that we strive to improve the future, we are committed to developing everyone who joins us on this journey, both personally and professionally.

Would you like to be part of laying the foundation for the way we live - and for the world around us?

Background of the assignment

The Analytics team at LKAB develops and maintains the company's analytics products, delivered as a portfolio of apps on the team's data platform. AI agents that build semantic models, reports and dashboards through MCP servers and command-line interfaces (CLIs) have quickly become practical tools in analytics development, both across the industry and in the team's own work.

The question is no longer whether agents are capable of building analytics products, but what they need to know to build them correctly: how key metrics are defined, which conventions and business rules to apply, and where the same term can mean different things in different parts of the business. Both research and industry experience suggest that it is this context, rather than the capability of the AI model itself, that is the main bottleneck. Yet there is little evidence, based on real data from large industrial companies like LKAB, on how context should be leveraged to create the most value from agentic development.

To address this, the Analytics team wants to build a context layer for its data platform: a structured, version-controlled collection of metric definitions, business rules, ontologies, conventions and verified examples that AI agents can draw on while they work. This thesis will design and build that context layer, and evaluate how much each part of it contributes to the quality of the agents' work.

Problem statement

The thesis addresses three research questions:

1. How much does context (such as metric definitions, business rules, ontologies, conventions and verified examples) improve the quality of agentic development, and what is the value of each new context block relative to the effort needed to produce it?

2. To what extent can a robust context layer narrow the gap between less and more capable AI models?

3. In a pull request-based workflow, what share of the agent's errors is caught by automated quality gates, and what share only by human review?

To answer these, you will build an agent environment where AI agents work against the team's BI platform and data warehouse through MCP servers and CLIs, with automated quality gates. You will also build an evaluation suite of real development tasks, drawn from both well-defined and ambiguous parts of the business, with existing validated reports as ground truth. The evaluation is set up as a controlled experiment that adds context incrementally, removes individual context types (ablation), compares a less and a more capable AI model, and repeats each run several times. The suite is designed to be re-run as AI models evolve, so that the findings stay relevant over time.

The context layer, agent environment and evaluation suite are meant to be implemented in the real workflow of the development team, and stay in use by the team after the thesis is completed.

What you bring

You are currently enrolled in a Master of Science in Engineering or a master’s program in computer science, computer engineering, data science, engineering physics, or a related field, with a focus on software development, data, or AI.

You have strong programming skills, are comfortable with Python and Git and have a working knowledge of SQL and data modeling. You have experience with, or a strong interest in, agentic development methodologies, AI agents, large language models and how to evaluate them. Experience with BI tools, cloud tools, data warehouses and modern data platforms is a plus but not a requirement.

You are self-motivated and curious, enjoy building things that work in the real world, and are comfortable designing and running your own experiments. You will work closely with the Analytics team and gain first-hand insight into how LKAB uses data and AI in practice.

Compensation for thesis projects and practical benefits at LKAB

For a thesis project approved by both the university and LKAB, we offer the following compensation:

• 35 000 SEK for a thesis equivalent to 30 ECTS credits.
• 20 000 SEK for a thesis equivalent to 15 ECTS credits.

If two students write together, the compensation is shared.

In addition to financial compensation, we offer:

• Travel reimbursement

  between the study location and LKAB

• Accommodation can be provided when physical presence on-site is required

More information

Start: January 2027

Location: Luleå/Malmberget/Kiruna. Alternatively, remote work with occasional on-site visits

Scope: 30 ECTS

For more information about the thesis project, you are welcome to contact the supervisor/mentor: Olof Hallqvist, [email protected]

If you have any questions regarding practical matters beyond the thesis project itself, please feel free to contact [email protected]

We appreciate your commitment and will get back to you if we consider your application suitable for the assignment. Thank you for wanting to do your thesis with us!