About

From particle physics to production systems

I trained as an experimental particle physicist. My doctorate at the University of Sussex was carried out on the ATLAS experiment at CERN, searching for supersymmetric particles in events with three leptons. The work was a lesson in looking for a very small signal inside an enormous amount of noisy data, and in being able to defend every step between the raw detector output and the final statement.

That turns out to describe a great deal of industrial machine learning as well. Since 2018 I have worked on the same shape of problem in industry: first as a Senior Research Scientist at Moogsoft, an AIOps company, where I worked on natural language processing for log classification. The classifier was already in place when I joined; my contribution was the domain-specific corpus that fed it, the ablation studies that showed what actually mattered, and the evaluation that let it be trusted in a production service. Then as Senior Advisor, Data Science at Dell Technologies, working on anomaly detection and forecasting for infrastructure at very large scale.

How I work

I prefer methods I can explain to the person who has to live with them. Complex models are sometimes necessary and often not, and the way to find out is to build the honest baseline first. I would rather tell a client that their existing approach is adequate than sell them a replacement.

I write my own code and I hand it over in a state someone else can maintain. Rust when the compute budget is real, Python when the work is exploratory. The documentation is part of the deliverable rather than an afterthought, because a system nobody can modify is a system with an expiry date.

I also think evaluation deserves more attention than it usually gets. Leakage, optimistic backtests and intervals that do not hold their coverage are the commonest ways for a project to look successful and be wrong. Checking for them is unglamorous and it is most of the job.

The same habits carry into expert witness work. When a dispute turns on how a model was built or whether a claim about it holds, the job is to read what was actually done, say what the evidence supports, and give the same answer whoever is asking. I take instructions from solicitors on matters involving machine learning systems, and I will decline a matter that is outside my expertise.

Research and writing

I have four first-author publications with the IEEE, covering log message classification and anomaly detection, and I co-chaired the AnNet workshop at IEEE/IFIP NOMS 2023. I have written about machine learning and statistics at researchdatapod.com since 2019, including an analysis of the scaling behaviour of GPT-3 published in August 2020.

• Google Scholar profile
• researchdatapod.com
• Selected writing

The company

Y. Shehu Consulting Ltd was incorporated in England and Wales in September 2026 and works with clients from the south coast of England and remotely. It is a one-person firm by design. When a piece of work needs more hands than I have, I will say so rather than quietly subcontract it.