What a software engineer misinterprets about device physics

Most of the last two years went into shipping. Klinn AI is production CRMs, a multi-tenant builder, SEO sites that have to rank. Before that: applied AI at OSCOWL in Hyderabad's deep-tech pocket, and bio-inspired drones at TiHAN. The master's at OVGU Magdeburg — Advanced Semiconductor Nanotechnologies — is the layer I kept walking around.

A software engineer can get very good at models and products without ever touching a device. You learn to talk about GPUs as rent, latency as a budget, and 'the hardware' as a constraint handed down from someone else. That is a reasonable way to ship. It is a poor way to understand why the bill looks like that.

Device physics is not a slower version of systems engineering. Fabrication and characterisation have their own failure modes. A process window is not an API. A wafer does not retry.

At OSCOWL the work sat across speech models, UAVs and custom silicon — research that had to survive contact with a product. The software instinct is to compress, quantise, batch, and cache. All of that is real. None of it changes the fact that you are still moving data to the arithmetic.

Neuromorphic and in-memory compute are the other end of the same problem: devices that do the arithmetic where the data already sits. I have spent two years on the software side of that cost. I want to be genuinely good at the measurement side — nanoscale device characterisation, not just the slides about it.

That is the doctoral work I am looking for: the boundary between device physics and the systems built on top of it. Not a career change. A missing layer.

Shipping an Edmissions World CRM teaches you about roles, status machines, and midnight jobs. Publishing on ornithopter navigation teaches you about estimators that have to respect a plant. Neither teaches you what a device is doing when a multiply happens.

The misinterpretation is not that software is shallow. It is that we treat the stack as if it ended at CUDA. Below that there is still a transistor, a dielectric, a contact, a measurement that may not agree with the model. OVGU is where I am supposed to stop hand-waving that.

I wrote the About page around this on purpose. Path, publication, interests — neuromorphic compute, nanoscale characterisation, autonomy under constraint. The writings exist so those claims have arguments, not just a kicker.

I am in Magdeburg because the programme is fabrication and characterisation, not another year of shipping. Models and products are downstream of devices. I already know how to ship. I do not yet know this layer well enough to be useful at it.

If you work in software and you feel the bill in your bones, that feeling is pointing at physics. You can ignore it and keep making products. I did, for two years. Then I booked the master's.