If you are interested in any of these projects, please email your CV and a brief transcript to Dr. Behzad Bozorgtabar at behzad@ece.au.dk.

1. Agentic AI for Scientific Discovery

• Motivation: Automated computer vision tools are fundamentally reshaping scientific discovery in biomedical imaging. While "Code-Writing Agents" can automate the adaptation of these tools, they face a critical limitation: Overfitting.

• Objective: Develop robust, "anti-overfitting" mechanisms for self-coding agents utilizing modern open-source agent frameworks.

• Requirements: Strong Python skills (OpenCV, skimage, PyTorch), knowledge of ML and Computer Vision, and familiarity with LLM API integration.

2. MAC-Health: Multi-Agent AI Copilot for Healthcare

• Motivation: Clinical diagnosis necessitates collaborative decision-making. Multi-Agent Systems address single-model hallucinations by assigning distinct roles to different models.

• Objective: Build the MAC-Health architecture using agentic frameworks (e.g., CrewAI or LangGraph) and implement Retrieval-Augmented Generation (RAG) to connect agents to medical guidelines.

• Requirements: Strong programming skills in Python, foundational knowledge of NLP and LLMs, and familiarity with API integration or RAG pipelines.

3. Parameter-Efficient Test-Time Adaptation for Earth Observation

• Motivation: Geospatial foundation models drop in performance when evaluated on out-of-distribution geographic domains. Standard Test-Time Training (TTT) mitigates this but is highly memory-intensive.

• Objective: Develop a Parameter-Efficient Test-Time Training (PE-TTT) framework to achieve geographic generalization while reducing GPU memory requirements.

• Requirements: Strong Python and PyTorch skills, understanding of Deep Learning (ViTs or CNNs), and familiarity with domain adaptation or LoRA.