Regardless of proposed project below, you will be involved in the process of further scoping and formulation of the final research question to make sure that the project fits your interests. That said, some of them are more fixed in scope from the beginning whereas some are more open ended.

1. Data compliance layer

   • The upcoming EU AI Act requires certain AI systems to document and trace their data usage. This project aims to develop a data compliance layer that transparently logs data consumption — including what data is accessed, when, and how often — to ensure full traceability of training and evaluation datasets. The goal is to measure the performance trade-off of doing this and implement tests that showcases that even under extreme data shuffling results can be reproduced.
   • https://digital-strategy.ec.europa.eu/en/faqs/guidelines-and-code-practice-transparent-ai-systems

2. Comparison of Deployment frameworks

   • Evaluate and compare the performance of different deployment frameworks (e.g., vllm, fastapi) for serving ML models. Analyze metrics like latency, throughput, and resource utilization to identify the optimal deployment solution for various use cases.
   • https://www.mlsysbook.ai/contents/core/benchmarking/benchmarking.html

3. SOTA monitoring methods

   • Investigate current approaches to detecting drift in production machine learning systems. Design and implement monitoring solutions that can identify when model performance degrades due to data distribution changes. Focus on developing practical criteria for determining optimal retraining schedules to balance computational costs with model performance.
   • https://arxiv.org/pdf/2203.11070

4. Quantization + pruning

   • Investigate current literature on the intersection of pruning and quantization for creating efficient neural networks. The main work consist of re-implementing the algorithm in the below paper and investigate it for other domains than what is investigated in the paper (other datasets, other models etc.)
   • https://openaccess.thecvf.com/content/CVPR2025/papers/Qu_Automatic_Joint_Structured_Pruning_and_Quantization_for_Efficient_Neural_Network_CVPR_2025_paper.pdf

5. RAG for DTU

   • DTU hosts a vast and diverse knowledge base that can be challenging for students to navigate. This project focuses on developing a Retrieval-Augmented Generation (RAG) system integrated with DTU’s locally hosted LLM infrastructure to provide accurate, context-aware answers to common student queries. The goal is to improve information accessibility while maintaining data privacy and sovereignty.
   • https://campusai.compute.dtu.dk

6. LLM for speech assessment (in collaboration with GN):

   • This project explores recent advances in Large Audio Models (LAMs) for descriptive speech quality assessment. The student will replicate and evaluate open-source LAMs (e.g., Qwen2-Audio, SALMONN) on public and internal audio datasets, with a focus on adapting them for efficient, sovereign (non-cloud) inference = Gefion Supercompute.
   • https://arxiv.org/pdf/2501.17202

7. Quantization methods for audio (in collaboration with GN)

   • Quantization is a common method for creating efficient neural networks. However, its impact on audio models remains understudied. This project investigates quantization techniques for speech and audio tasks, benchmarking different strategies (e.g., post-training vs. quantization-aware) on open and internal datasets, with a focus on maintaining perceptual quality and model efficiency.

8. Models for dementia detection (in collaboration with DemensAI)

   • Develop models for early detection of dementia from speech and language data. The project involves exploring multimodal or speech-based machine learning approaches, benchmarking existing open-source models, and adapting them for clinical or real-world screening scenarios. Focus may include model interpretability, data efficiency, or robustness to recording conditions.
   • https://www.demensai.dk

9. Quality control using vision (in collaboration with JLI)

   • This project develops and optimizes deep learning vision systems (e.g., CNNs, Vision Transformers) for automated quality control (AQC) in manufacturing. It involves benchmarking models for defect detection and classification using limited data (transfer learning, few-shot methods) because annotation is often a very manual and labor intensive process.
   • Project is not fully specified because JLIs availability differs from semester to semester, if interested let me know and I reach out to them

10. LLM as a judge

    • Current approach to evaluating output of LLMs matches the expected output involves another LLM, often called an LLM-as-a-judge. There are a lot of open-questions in how to proper set up LLM-as-a-judge systems. This project aims to investigate one or more and setup a proper system for doing so.
    • https://arxiv.org/pdf/2411.15594

11. Inverse HealthBench (in collaboration with Laerdal Medical)

    • Creation of a patient-centered benchmark and dataset capturing how real patients describe symptoms, ask questions, and react to clinical advice (text + optional short audio). This project collects diverse, de-identified patient utterances across ages, languages and settings, and annotates them for intent, urgency, misinformation, emotional tone, health-literacy and actionability so models and clinicians can be evaluated and trained on authentic patient-facing behaviour.
    • https://openai.com/da-DK/index/healthbench/

12. Spiking neural networks (in collaboration with MicroChip)

    • Design and implement a low-power neuromorphic inference framework for edge devices using Spiking Neural Networks (SNNs). The project adapts the open-source SpikerPlus framework to Microchip MPUs or FPGAs, evaluating trade-offs between accuracy, latency, and energy efficiency. The outcome is a proof-of-concept demonstrating real-time, energy-efficient intelligence for resource-constrained applications like drones, robotics, and medical devices.

      SNN-ProjectProposal.pdf

13. ???

14. ???