Master Thesis in Software Engineering: Adaptive Architecture for Human-AI Annotation Systems

Location: Göteborg, SE, 417 15

Master Thesis in Software Engineering: Adaptive Architecture for Human-AI Annotation Systems

Location

Göteborg, SE, 417 15

#job-location.job-location-inline {

display: inline;

}

Position Type

Student

Permanent/Temporary

Temporary

Transport is at the core of modern society. Imagine using your expertise to shape sustainable transport and infrastructure solutions for the future. If you seek to make a difference on a global scale, working with next-gen technologies and the sharpest collaborative teams, then we could be a perfect match.

Thesis Scope & Problem Statement

Volvo Autonomous Solutions develops systems for automated hauling operations in quarries and open pit mines using Volvo trucks. Validating these systems requires high-quality ground-truth datasets. Our team owns a camera-based annotation tool that produces bounding boxes used as a reference for validating LiDAR- and radar-based moving-object perception, with camera perception entering the stack. Errors in these reference annotations can compromise the validity of downstream evaluations.

A previous Master's thesis evaluated different architectures for combining AI-generated annotations wi human verification. The results highlighted trade-offs between annotation correctness, proces performance, maintainability, and human effort: no single static configuration performed best across these dimensions. Participants also reported verifying less actively when AI suggestions appeared plausible, highlighting a potential risk of automation bias.

This thesis project is intended for two Master's students working together.

The central focus of this Master's thesis is to design, implement, and evaluate a confidence-aware adapti verification architecture within the existing annotation tool. Instead of applying the same review process every detection or scene, the system will use confidence signals to select an appropriate depth of hum verification. This is a software engineering thesis focused on how to structure, integrate, test, and evalua that adaptive behavior, with particular attention to maintainability, traceability, fallback behavior, and effective human oversight. Existing AI models provide the application context; developing new percepti models is not the primary objective.

Research Objectives & Tasks

The work may include

Design and integrate a modular architecture with explicit interfaces for confidence provider verification policies, decision logging, and annotator feedbac

Implement confidence-provider components that expose calibrated or normalized signals, such model confidence, model agreement, source model, and object class frequency, through a comm interface.

Implement and compare interchangeable threshold-based and learned decision policies for routing cases to light review, full review, or a second review. Automatic acceptance will be considered only where permitted by safety policy, with low-confidence, out-of-distribution, and safety-critical cas routed to human review.

Integrate attention cues into the annotation interface and record the signals, policy version, selected review mode, and human outcome for each routing decision.

Develop automated tests for routing decisions, component interfaces, and fallback behavior.

Benchmark the adaptive architecture against existing manual, sequential, and parallel workflo using annotation correctness, recall on safety-critical classes, error catch rates, task completion time, routing overhead, and human intervention.

Evaluate maintainability and testability through change scenarios, such as adding a confiden provider or replacing a decision policy.

Conduct a counterbalanced user study to investigate verification behavior, cognitive workload, tru calibration, and susceptibility to automation bias.

Depending on feasibility and measurable added value, the work may also explore out-of-distribution detection or vision-language-model-derived scene complexity as additional confidence signals. The co evaluation will use fixed policies during controlled experiments; online adaptation is an optional extension.

Who Are We?

You will join the Verification & Validation (V&V) Virtual Driver team at V.A.S., focused on ensuring t integrity of the software stack for autonomous decision-making and control. The team brings together developers with varied backgrounds and seniority, all centered around robust V&V practices. The environment is friendly and helpful, with both in-office development and in-vehicle testing.

Who Are You?

You are a final-year Master's student in Software Engineering, Computer Science, or a related field, with strong interest in software architecture, software quality, and building maintainable systems. You are comfortable programming in Python and interested in extending and evaluating an existing software platform in an industrial setting.

You should have a foundation in modular software design, interface design, automated testing, and version control, together with an interest in empirical software engineering. You should be motivated to compare design alternatives and connect architectural decisions with measurable effects on correctnes performance, maintainability, and testability.

Familiarity with AI-assisted systems, machine learning, computer vision, or human-computer interaction is advantageous, but prior specialization in machine learning or object detection is not required. An interest in learning enough about model confidence and human verification to make sound software design decis is important.

Why Join Us?

This thesis offers an opportunity to apply software engineering methods to a concrete industrial challeng designing an adaptive human-AI workflow that is maintainable, testable, and auditable. You will build on existing annotation platform and prior thesis results, taking the work from architectural design through implementation and empirical evaluation.

You will

Gain hands-on experience in a