Bachelor and Master Theses

Engineers maintaining a system need to understand why earlier architectural decisions were made. AI can produce convincing explanations, but these may differ from the reasons documented by the original developers. This thesis investigates whether AI can recover evidence-supported explanations and recognize when the available records are insufficient.

Possible Research Goal

Evaluate AI’s ability to explain architectural decisions while distinguishing documented reasons from unsupported interpretations.

Possible  Research Questions

1. How accurately can AI identify the documented reasons behind architectural decisions?
2. Does requiring source-linked evidence improve accuracy and reduce unsupported explanations?
3. How do incomplete records and later decision revisions affect the results?

Possible Research Method

Controlled experiment using public GitHub repositories:

1. Select projects containing architectural decision records, design proposals and linked discussions.
2. Construct a validated dataset of decisions, documented reasons and supporting evidence.
3. Compare LLMs 
4. Repeat the evaluation with missing evidence and genuine later revisions.
5. Measure supported-reason recovery, unsupported claims, evidence accuracy and execution cost.

The initial target is 120 decisions across 10–12 projects, refined through a pilot.

Programming and scripting, preferably Python.

Git and GitHub, including repository history and API access.

Automated data collection, processing and experiment execution.

Working with structured data and testing data-processing pipelines.

Knowledge of software architecture is required.

Experience with LLM APIs.