Dimitar Gueorguiev

I reformulated the algorithm as a mixed-integer problem by linearizing the original objective function, resulting in a multifold speedup of the core algorithm. I redesigned the fulfillment algorithm using the ε-constraint method to replace the weighted objectives method for scalarization. Additionally, I researched various reinforcement learning policy gradient algorithms, such as PPO, to potentially replace the action-value reward-based reinforcement learning component of the algorithm.

I developed a new frequency domain forecaster that predicts future trends and seasonal or cyclical events in retail and fashion industry datasets. To achieve this, I created a super-resolution algorithm that formulates the problem as a convex optimization task in the frequency domain and solves it using a fast Fourier transform (FFT) relaxation technique.

I developed a collection of algorithms for smooth gradient outpainting. This unfinished work was created over two months to meet Nike Digital Marketing's need to add outpainting regions with smooth gradient backgrounds to Nike images. These algorithms enable the marketing team to generate a much wider variety of image crops across a large set of aspect ratios. I plan to complete this project in the near future.

I developed WorkloadPlanner, a simulation module integrated into the Unisphere for the VMAX software product. I designed and implemented the component simulation models for central processing units (physical and logical cores), InfiniBand fabric, I/O boards, disks, and I/O ports, all developed and validated in Java. WorkloadPlanner simulates the performance of the target array by calculating component utilizations and back-end response times.

Additionally, I created the WorkloadPlanner FAST model, a constrained optimization solver written in Java. This model determines whether incoming workloads can be accommodated on existing target array pools while respecting performance and capacity constraints. It ensures that storage group SLOs and storage pool SLE response times are not violated and estimates the available headroom on the storage array back end.