Master’s Thesis – Context Compilation for Trustworthy AI

Job Description:

Engineering repositories hold typed, versioned, traceable models. Language models usually receive a flattened document or a handful of retrieved fragments, losing the dependency that decides the answer, and treating an assumption as a verified fact.

This thesis asks whether context should be compiled rather than retrieved. A deterministic rule engine derives what a question actually depends on, keeps a record of exactly which facts the derivation used, and hands the language model that record. The context is then sufficient by construction, which makes the interesting question whether the model actually uses what it was given.

You will build the prototype against a real industrial model, design the evaluation yourself, and test whether stating what is verified, assumed, stale or unknown measurably changes how reliably an LLM reasons. Hosted at our Gothenburg headquarters and supervised jointly with your university.

Start: January 2027 · 30 ECTS, 20 weeks · Open to a pair, or a single exceptional candidate

Key Responsibilities:

• Build a typed model adapter and a rule-based closure engine with provenance tracking

• Design a question set and an evaluation protocol with expert-reviewed answers

• Compare compiled context against document and graph retrieval under matched context budgets

• Measure whether explicit evidence status reduces unsupported claims

• Document limitations, failure modes and negative results

Requirements:

• Enrolled in an MSc in computer science, data science, engineering physics, systems engineering or similar, with the thesis remaining

• Strong programming, and the judgement to build a small system rather than a script

• Comfortable with graphs and typed data as a way of thinking: you have written a traversal, a query planner, a constraint or type checker

• Working knowledge of modern LLM tooling: prompting, structured output, building an evaluation harness

• Enough statistical literacy to design a fair comparison and not over-read a result

• Precision in writing: a core contribution of this thesis is a definition

• Must be located within commutable distance of Gothenburg, Sweden

Meriting, not required:

• Logic programming, Datalog, Prolog, knowledge graphs or formal methods

• Experience in a safety-critical or regulated domain

• MBSE, SysML or systems engineering exposure

No prior knowledge of SystemWeaver, MBSE or automotive safety standards is expected. We will teach you what you need.

Required Skills:

Python or C#, graph algorithms, typed data models, LLM evaluation, prompt engineering, structured output, experimental design, applied statistics, technical writing, logic programming (meriting), knowledge graphs (meriting), formal methods (meriting).

What you get:

• A real industrial engineering model to work on, not a benchmark dataset

• A sharply scoped research question and results worth publishing

• Modern workspace at CityGate, Gothenburg’s tallest commercial building

• Flexible hours

Applying:

Interviews are conducted on a rolling basis, and the position will be filled as soon as we find the right fit. So, apply early rather than waiting for a deadline!

Applications in pairs are preferred, but if you happen to be a unicorn who can do it all, we’d love to hear from you too! Just mention it in your cover letter. Otherwise, find your partner and apply together.

Contact: Shahid Muhammad, shahid.muhammad@systemweaver.com