Experience

A few pieces of work that show the shape of what I do. Details that belong to former employers and their customers are left out, so these are deliberately generic in places.

Anomaly detection for a 9,000-GPU supercomputing fleet

Situation: A leading US national laboratory operated a supercomputing fleet of roughly 9,000 GPUs. Faults on individual devices were expensive in wasted allocation and hard to find early, and the volume of telemetry made manual inspection impossible.
Work: I designed and built a production anomaly-detection engine for the fleet, covering the whole path from method selection through implementation to a deployable service. The engine was written for the compute budget it had to run in, and it came with a validation harness so that its behaviour could be reproduced and challenged rather than taken on trust.
Outcome: A working detection service running against live fleet telemetry, documented to the point where other engineers could take it on, with the limitations of the method written down alongside its results.

Work performed while employed at Dell Technologies.

Workload forecasting and what-if planning

Situation: Product and infrastructure teams needed to answer capacity questions that were being handled by intuition and spreadsheets: what happens to this system if the workload changes, and how much headroom is actually left.
Work: I built a workload forecasting and what-if planning tool that non-specialists could drive themselves. Most of the effort went into the evaluation rather than the model: finding and closing leakage, checking that the intervals held their stated coverage, and making the tool explain why an answer had changed.
Outcome: The tool was adopted by product teams and used for planning decisions, replacing a manual process with one whose assumptions were visible.

Work performed while employed at Dell Technologies.

Log message classification, from research to a shipped service

Situation: An AIOps platform needed to make sense of unstructured log messages arriving from many different systems, at a volume and variety that ruled out hand-written rules. A classifier existed. What was missing was the evidence that it could be trusted, and the domain knowledge to make it better.
Work: I built the domain-specific corpus that fed the model, crawling vendor documentation to capture the vocabulary of real infrastructure logs, and ran the ablation studies that showed which parts of the approach were actually doing the work. I designed the evaluation and published the results as first author with the IEEE, reaching 0.98 accuracy on log message classification.
Outcome: Peer-reviewed results, and a running service rather than a prototype. The evaluation work was what let the system be relied on in production.

Work performed while employed at Moogsoft.

Searching for new particles at CERN

Situation: The ATLAS experiment records collisions at a rate no person could ever inspect. Any new signal would be vanishingly rare and would sit underneath a very large, well-understood background.
Work: My doctoral research searched for the electroweak production of supersymmetric particles in events containing three leptons, in the difficult region where the masses involved are close together and the signal is easiest to lose.
Outcome: A completed doctorate, and habits that apply as much in industry as in research: treat the background properly, quantify the uncertainty, and never report a result you cannot reproduce.

Doctoral research, University of Sussex and the ATLAS collaboration.

A similar problem, or a different one

If your situation looks like one of these, the conversation is easy. If it does not, describe it anyway.

[email protected]