Machine Learning

Supervisors: Faizan Ahmed(UT), Pieter Zeilstra (Saxion University of Applied Sciences)

Be part of an international research group working on the Horizon 2020 SUSTENANCE project, where the goal is to help local communities use renewable energy like solar and wind more efficiently and become less dependent on the main power grid.

Your assignment will be about extending an existing platform called IECON. This platform runs on small computers like Raspberry Pi and collects real-time energy data from households. Right now, this data stays mostly within the home. Your challenge is to make it possible for households to share their data securely and with full control, so they can decide who sees their data and how it is used.

To do this, you will explore the concept of data spaces, which are secure digital environments where data can be shared based on clear rules and privacy protections. You will research existing tools such as IDS connectors, talk to stakeholders to understand what they need, and then design, build, and test your own solution that fits on edge devices like Raspberry Pi.

If you are interested in sustainability, data privacy, and working with real-world technologies that have a global impact, this project offers a chance to make a real difference while building valuable skills for your future.

Supervisors: Milan Lopuhaä-Zwakenberg, Benedikt Peterseim (contact: benedikt.peterseim@utwente.nl)

Large language models (LLMs), particularly recent advances in so-called reasoning models, demonstrate remarkable performance in classical computer algebra tasks such as symbolic integration. However, they remain unreliable—sometimes showing what is referred to as “hallucinations”—without any formal guarantee of correctness. This limits their practical usefulness, as manually verifying LLM-generated solutions can be cumbersome. The issue is exacerbated by the fact that incorrect outputs may appear logical and convincing at first glance due to the nature of LLM training.

This Master’s thesis project aims to integrate LLMs with formal methods to ensure correctness in specific, well-defined computer algebra tasks.

This may include, but will not be limited to:

• Designing a domain-specific formal grammar that the LLM must adhere to when outputting its final result.
• Applying grammar-constrained decoding [1] to force the LMM to provide its final answer in this specified formal language. 
• Utilising external verifiers to check the parsed results or, if necessary and sufficiently simple, developing simple verification tools for this purpose.
• Leveraging LLMs to translate informal user input into the specified formal language. This presents a creative challenge: the formal language should be close enough to standard calculus notation to allow users to intuitively assess the correctness of the translation.
• Evaluating this approach by comparing its performance to existing computer algebra systems.

[1] Scholak, Torsten, Nathan Schucher, and Dzmitry Bahdanau. "PICARD: Parsing incrementally for constrained auto-regressive decoding from language models." Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, November 2021.

Supervisors: Faizan Ahmed

Point clouds are sets of spatial data points captured by 3D scanning techniques such as lidar. These point clouds contain many million points of data, resulting in 3D representations of the railway environment. Point cloud data can be used to create machine learning models that can classify the object in rail infrastructure automatically. 

The main research question is: How to reliably segment, classify, and visualize point clouds for railway catenary systems with scant computational power, memory, and ground truth labels?

However, the student can focus on a combination of topics given below:

• leverage existing trained models for point cloud labeling
• pre-process (down-sampling. Augmentation, filtering, voxelization etc.) point clouds for accurate model training
• train deep learning models for point cloud segmentation and object detection
• analyze the variation in object sizes with relation to the accuracy of the trained models.
• analyze the effect of point cloud density on accuracy
• determine the accuracy of the trained model objectively (or What are the excellent metric for such machine learning tasks?)
• connecting point clouds with existing 3D object libraries

UT Supervisor: Faizan Ahmed, Saxion Supervisors:  Bram Ton, Selin Çolakhasanoglu

There has been an explosive use of Large Language Models (LLM) for all kinds of applications. One of these applications is a RAG system. In essence, it allows users to query their own documents. For instance, for a building services engineer, he/she can upload all manuals for the products he/she works on and can easily query them. An example query could be: “Today I am working on a Remeha Calante 40c central heating unit and the error on the screen says E14. What does it mean, and how do I resolve the underlying issue?”

These RAG systems work well in practice, but we do not have objective ways of evaluating them. One method of evaluation would be to create a golden data set of possible question answer pairs and measure how well a RAG system answers them. Creating these golden datasets by hand would be too costly, and would require a new dataset per application domain.

Therefore, we want to explore the automatic generation of good quality Q-A pairs using LLM’s. To judge the quality of the Q-pairs we want to use a …. LLM. Within this loop we can apply prompt tuning to optimize the quality of the Q-A pairs. This assignment is ideal for students interested in natural language processing, evaluation methodologies, and practical AI applications in real-world technical domains like engineering and building systems. You will gain hands-on experience with LLM workflows while contributing to research that helps professionals make better, faster decisions on the job.

Task Description

• Use existing Python frameworks (such as LangChain) to create this workflow.
• Define criteria for a good Question-Answer pair
• (Optionally) compare user evaluation vs automated evaluation

Practical Information

• Student profile: Bachelor/Master internship
• Contact Person UT: Faizan Ahmed
• Contact person(s) at Saxion:  Bram Ton (b.t.ton@saxion.nl), Selin Çolakhasanoglu (s.colakhasanoglu@saxion.nl)
• Lectoraat Ambient Intelligence: saxion.nl/ami