PulpaTronics
London
I spent my placement at PulpaTronics, a startup developing RFID tags made from paper instead of plastic and metal. RFID tags are the small labels that let a shop scan an item without touching it, and billions are thrown away every year. Making them from paper removes a large amount of waste from the supply chain, but it also makes them much harder to design: paper behaves less predictably than metal, so a design that works on a computer does not always work on a bench. Designing one of these tags means testing an antenna shape, and every test is slow. A single simulation ties up a powerful computer for minutes, and measuring the printed material by hand takes an engineer away from other work. With thousands of possible shapes, testing them one at a time is simply not practical. My job was to remove those bottlenecks with software. I built systems that learn from previous simulations and measurements to help identify promising designs without exhaustively testing every possibility. This allows the team to explore much larger design spaces while focusing expensive simulations and laboratory work where they are most valuable.
I worked in the computational engineering team, with responsibility for three projects. Each removes a different bottleneck from the design process, and all three run on cloud computing that I set up and costed myself. The first and largest was an automated design system for antennas. Rather than an engineer proposing shapes one at a time, the system proposes them itself: it learns from the simulations it has already run, predicts which untested shapes look most promising, and spends expensive computing time only on those. It now runs automatically overnight, compares numerous families of antenna shapes, and ranks the best performers. The second replaced a manual laboratory measurement with automated processes using artificial intelligence. I assembled the equipment that photographs printed material and trained a machine learning model to estimate its electrical properties directly from the images, so a measurement that previously required hands-on work by technical experts can now be made by anyone from a photograph. The third, uses real-world measurements taken in a specialist testing environment to correct the simulations themselves, closing the gap between what the computer predicts and what the laboratory records.
The placement sat at the meeting point of three fields I had previously studied separately: electromagnetics, machine learning and software engineering. Bringing them together was the real work. I already knew how to build machine learning and computer vision models. What was new was applying them to a physical production-grade system where the answer is not simply a number that scores well but a design that has to behave correctly once it is manufactured and tested in a laboratory. Alongside this I applied antenna theory from my degree, and learned to treat simulation as something a program drives automatically thousands of times rather than something a person runs by hand. I set up and costed the cloud computing needed to do that at scale, which was entirely new to me. The largest shift was in how I write software. Automated testing, documentation and careful version control matter far more when a system runs unattended overnight and colleagues depend on its results.
Working in a small company gave me huge responsibility and ownership over my work, with degrees of freedom I had not been used to. I owned three projects at once, and embarked on projects outside my initial scope. Managing that meant sequencing the projects around their dependencies. It also meant talking to people well outside my own domain. I gave a session on large language models to colleagues without a machine learning background, interviewed candidates for the team, and spoke to investors and partners about what we were building. Each of these needed a different approach, and none of them was something I had done as a student. I learned to say early when something is not working. I spent my first weeks trying to run a simulation tool on cloud computers in a way it was never designed to work. The instinct is to keep going quietly until it does, but surfacing it early meant the team could rethink the approach together rather than after the fact.