Implement hardware-dependent knowledge distillation for a compact multispectral object-detection student, then run the compiled student on commercially available M.2 AI accelerators that can be plugged into our DPUs. In parallel, perform a trade-off study on knowledge distillation methods suited to this task.

The internship lasts 3 to 6 months and can start in January or February 2027.

•   Train a compact multispectral object-detection student from a teacher model, with Hinton-style knowledge distillation as the baseline.

•   Research and implement other knowledge distillation methods, and define a metric for how efficient the distillation process is.

•   Compile and run your student on two M.2 accelerators: Axelera Embedded 110m and MemryX Cascade 100.

•   Compare accuracy and throughput, assess toolchain usability, and recommend which hardware to use.

•   A bachelor or master student in computer science or AI.

•   Experienced in, or interested in, resource-constrained edge AI systems.

•   Curious about model compression and edge deployment.

•   Eager to learn, even if you have not worked with knowledge distillation before.

•   Hands-on experience with real edge AI hardware: two commercial edge AI accelerators and their software toolchains.

•   Practical skills in knowledge distillation, quantisation and model compilation for edge deployment.

•   A first-hand look at AI for space: fitting models into the tight power and compute budgets of onboard satellite processing.

•   Exposure to multispectral Earth observation data, which is less common than standard RGB imagery.

•   Join a fast-growing spacetech startup with a young, driven team and a dynamic, hands-on culture.