Qualcomm
Cork, Ireland
From March to September 2026 I worked at Qualcomm's R&D site in Cork, Ireland, as part of the group that develops the software stack enabling AI models to run efficiently on the company's tensor processing unit. Modern LLMs are typically trained and have inference run in datacentres. Deploying those same models on an edge device introduces much tighter constraints on compute, memory, latency and energy consumption, and closing that gap is the purpose of the compiler and optimisation software the team builds: it translates models into representations suitable for specialised AI hardware and applies techniques such as post-training quantization to reduce their resource requirements. Across the six months my work moved from how models are translated for the hardware to how the results of that translation are measured, which gave me a view of the toolchain from both ends. The experience has given my final year at Imperial a clear direction. I now know that I want to work at the boundary between AI software and the hardware on which it runs, and my final-year module choices and project will build directly on the experience I gained in model compilation, quantization and the deployment of machine-learning systems on resource-constrained hardware.
My first project extended compiler support for newer open-weight LLMs available through platforms such as Hugging Face. These models increasingly use different architectural patterns, attention mechanisms and tensor operations, so the compiler must support a continually expanding set of graph structures and operators. I worked on the model-conversion pipeline, helping map high-level PyTorch operations to hardware-specific implementations. Day to day this meant tracing transformations through the compiler, investigating operations that were not yet supported, and confirming that the translated graphs produced numerically consistent results. My second project addressed model quality after compilation and quantization. I developed a new method for measuring how closely the transformed model matched the original model's behaviour. The resulting tool was adopted by the wider team and integrated into the automated checks that run whenever changes are made to the compiler software. This made it possible to catch accuracy regressions earlier, and to assess the effects of individual model transformations far more systematically than had previously been practical.
I came away with a working understanding of how a trained model is lowered onto specialised AI hardware, and of the trade-offs that edge deployment forces. Practically, I can now read a model at the graph level, follow an operation from its high-level PyTorch form through to a hardware-specific implementation, and diagnose why a given operator fails to convert. I learned both how post-training quantization reduces a model's resource footprint and why it can change behaviour in ways that are not obvious from the model definition alone. That led into the skill I value most from the placement: designing evaluation. Deciding what "close enough to the original" should mean, and building something rigorous enough that other engineers would trust it, taught me more about numerical validation than any coursework had. Finally, I learned to navigate a large industrial codebase, and to write tooling that fits an automated development process rather than sitting outside it.
Facing unfamiliar problems in a mature codebase taught me to decompose them, scope tasks realistically, validate changes carefully, and ask for help early rather than losing days to something a colleague could resolve in minutes. I also learned the value of taking initiative. After coming across a research paper written by another Qualcomm team on a problem related to mine, I arranged a meeting between the two groups so we could compare approaches. The discussion was productive on its own terms, but it also showed me how much is gained by deliberately connecting work across different parts of a large engineering organisation. Supporting a newer intern as they settled into the team developed my mentoring skills, and entering the site's intern poster competition with them required presenting technical work to a general audience; we won joint first place among roughly fifty interns. Working with colleagues in Ireland, the United States and the Middle East stretched my communication further still. Living independently in Cork, alongside interns from across Europe, made the placement as rewarding culturally as it was technically.