Machine learning applied to quantum computing.
I am a Senior Ad Tech Consultant at Amazon Advertising in New York, which is a technical solutions consulting role. In practice that means building and shipping production agentic systems: multi-tool LLM architectures with governance, human-in-the-loop control, and the evaluation harnesses that gate them before anything reaches a customer. Six years at Amazon across Bogotá, Munich and New York.
Alongside that I am doing a PhD at Universitat Oberta de Catalunya, in the SOM Research Lab, on ontology-driven verification of AI-assisted software and data pipelines. The thread running through it is the same one that runs through my day job and through this site: how do you know a system does what you think it does, and what does the evidence actually have to look like.
Machine learning applied to quantum computing. Classical deep learning is the method, quantum computing is the domain, and none of it runs on a quantum computer.
That framing is deliberate. Quantum machine learning, meaning learning algorithms executed on quantum hardware, mostly does not beat classical baselines at current scales, and the honest benchmarks say so. The direction that works today is the reverse: classical models solving quantum problems, where the data is free, the ground truth is exact, and the results are positive. Error correction decoding, device calibration, control optimisation, many-body simulation.
I am also finishing an MSc in Quantum Technology Applications at the University of Sussex, so the domain is not arbitrary.
Two habits show up in everything here and they are the ones I would want judged.
The first is that a baseline gets validated before anything is built on it. Every project starts by reproducing a number someone else already published. If the harness cannot recover a known result, nothing computed afterwards means anything, no matter how good the model looks.
The second is reporting what did not work with the same weight as what did. The decoder study has a section on the configurations where the models lost, and why, because the reason turned out to be my training budget rather than the architectures. Finding that took an extra experiment and changed the conclusion. Most write-ups quietly drop that section.
GitHub • Hugging Face • Google Scholar • ORCID • LinkedIn
valencia.diego825@gmail.com