Hey, I’m
Carlos.

Chips are built in layers, and those layers need to line up very precisely. If the displacement is too large, the part may have to be discarded. My work was about predicting that displacement and how uncertain the prediction was.

I used time-series models and statistical methods to predict overlay, and conformal prediction to quantify the uncertainty. The idea was to help the team anticipate alignment problems and know how much confidence to put in a prediction.

DHL Express needs to reserve airport capacity before the packages arrive. If more arrive than expected, extra space costs more. If fewer arrive, the reserved space goes unused.

I worked on predicting those volumes using graph neural networks, attention mechanisms and a new architecture we developed for the problem. The architecture was a variant of an augmented message-passing approach, similar in spirit to the ideas discussed in the paper below. Better forecasts would help DHL reserve the right amount of space and avoid those extra costs.

DHL Express has its own aircraft to move urgent shipments around the world. Planning their schedules means coordinating the whole network while accounting for the constraints of the operation.

I worked on a strategic scheduling problem that hadn’t been solved internally. It took a lot of research and conversations with people across the company to understand the problem and work out how to model it in a way that could scale to such a large network, with so many variables. I used a branch-and-price approach to break the problem into manageable pieces. The aim was to help planners coordinate the fleet more effectively.