Model evaluation. Data annotation. Quality that holds.

I work on the part of AI that decides whether a model is any good. Six years of it: reading generated output against a rubric, deciding what counts as correct, and recording the judgment consistently enough that thousands of them add up to a usable training signal.

At PhotoMath that meant 3,000 LaTeX-formatted mathematics problems written to train reasoning, and 15,000 generated solutions reviewed against scoring guidelines at 98.7% accuracy, which earned a Special Tasks Expert award. At Populii AI (Tech Mahindra) it broadened into multimodal work: bounding boxes, ROI extraction, segmentation and clone matching on image-to-math data, plus PII detection and content filtering against safety guidelines. Through Mercor I take on response ranking and dataset refinement across whatever domain the workflow needs.

Underneath is an electrical engineering degree and two years on high-voltage transmission infrastructure in Pakistan — testing, fault analysis, PLC and HMI control systems. That background is why I'm now finishing an MSc in Sustainability at Ulster University, modelling community energy systems in Northern Ireland. Rubrics and load-flow studies turn out to reward the same habit: define the criteria before you start, then apply them the same way every time.