Releasing a software product often is seen as the end point of the software development process: everything's
implemented, time to earn some cash! Of course, quite the opposite is true, since 40 to 80% of software
development costs are spent during software maintenance activities after release, i.e., to fix bugs,
improve the user experience and add new functionality. In the age of AI, maintenance also involves monitoring ML
models for performance degradation (drift) and managing technical debt in complex engineering pipelines.
MCIS' mission is to help practitioners maintain their AI-powered software systems.
For example, how can we detect and mitigate performance degradation (drift) in ML models? How do we manage
technical debt in machine learning pipelines? How can we coordinate vulnerability fixes across large-scale
software ecosystems? Which AI-product release-readiness checklists should be followed before deployment?
By empirical research on software development process data stored in revision control systems (Git, ...), mailing
list archives, bug repositories (Jira, GitHub Issues, ...), and modern model registries like Hugging Face, MCIS
addresses these challenges. Typically, the outcome of our research are models developed using data mining,
statistical analysis, and advanced machine learning techniques.
Turning textual source code into an installable software product, ready to execute on the user's hardware
platform, is at the heart of software construction. Practitioners need help releasing their traditional and
AI-powered products faster using optimized CI/CD and Infrastructure-as-Code, significantly shortening commit
turn-around time.
MCIS' mission is to help practitioners release their products faster, but without
sacrificing quality. To achieve this, one needs insight into the modern release engineering
process. For example, how can we reduce redundant continuous integration activity through commit
grouping and skip prediction? How do we optimize build batching algorithms at scale? How healthy is our software
supply chain, and why do some builds fail to be reproducible across different ecosystems? How can we prepare our
development process for seamless continuous delivery?
Using the same empirical techniques as for software maintenance, we address the questions above. In addition, our
lab has unique expertise and data sets on release engineering and build systems (Bazel, GitHub Actions, ...),
and close contacts with industry for access to "real-life" data.
Software Intelligence forms the heart of MCIS' research. We leverage Large Language Models (LLMs) and software
analytics to facilitate program comprehension and development. From boosting LLM-based code translation using
transitive intermediate translations to predicting code review hot-spots, we turn software data into actionable
developer intelligence.
MCIS leverages AI to help practitioners understand and develop their software systems and
infrastructure to facilitate maintenance, construction and other development activities. For example,
how can we enhance LLM-based code translation using transitive intermediate translations? Which files in a pull
request are most likely to need comments from code reviewers? How do we effectively manage ML assets and
navigate foundation model leaderboards?
Intelligence covers elements from advanced model versioning (Semantic Versioning on Hugging Face), large language
models for code, and human-AI collaboration. Given that knowledge is crucial in software development, software
intelligence is central to MCIS' mission.
