Osama Abdelaal

2 linked episodes

I work on making neural networks run on memory that behaves like physics rather than like logic. At Fraunhofer IPMS I characterize ferroelectric FeFET and FeMFET devices, then build the training methods that let real networks tolerate what those devices actually do. Analog compute-in-memory is fast and efficient. It is also noisy, drifty, and limited in precision. Most of the interesting work lives in the gap between what a device can physically deliver and what a network needs to stay accurate. My PhD at TU Braunschweig, supervised by Prof. Dr. Thomas Kämpfe, sits in that gap: device and algorithm co-design for FeFET compute-in-memory for edge neural inference in agentic applications. One idea I keep coming back to is that device noise does not have to be treated as damage. Bit errors from an imperfect memory cell start to look like a regularizer if you train with them instead of against them. The agentic half of that topic is not theoretical for me. Alongside the research I design and run production agentic AI systems: multi-agent LLM architectures handling autonomous sales, marketing, and customer service workflows against real users, with the orchestration, context, tooling, and evaluation layers that keep them standing up under load. Building those systems is what convinced me that inference cost and latency are the binding constraint on where agents can realistically be deployed. That is the demand side of the same problem the hardware work approaches from the supply side. Before that I spent several years on the applied end of edge inference. Model compression and quantization-aware training for low-precision hardware at Gemesys. Transformer-based vision systems running in real time on edge GPUs for automotive manufacturing at Unicontrol. Multi-agent reinforcement learning for coordinated industrial machines with TRUMPF. The path here was not linear. Chemistry, then computational structural modeling, then automotive software engineering, then production deep learning, now neuromorphic hardware. What connects them is physics-constrained computation: building models that have to respect what the material or the silicon underneath them can actually do. I also tutor the neuromorphic systems class and supervise student projects. Happy to talk about compute-in-memory, spiking networks, agentic systems, and inference at the edge.