Portfolio
Software Engineer
Trained in computer science, I initially specialized in operations research. I then went deep into artificial intelligence, building a broad foundation from classical modeling to generative methods. My wide-ranging curiosity drives me to learn new material thoroughly and help deliver working software. Recently, at Ghent University, I advanced diffusion models for generative antibody design using negative guidance. This work combines my interests in biology and AI. I think visually, build clean interfaces, and ship research-grade systems with measurable impact.
hello@simonvermeir.com
Optimizing Diffusion Models For Generative Antibody Design
ARTIFICIAL INTELLIGENCE
Negative Guidance with an auxiliary model to improve sample quality without increasing dataset size.
Started from existing model. Introduced guidance technique from image diffusion. Improved sampling strategy of original model. Retrained base model. Trained aux model using synthetic data from the base model. Via custom dataloader, uses saved tensors stored in LMDB (memory mapped for speed). Guidance on base model using aux model. Compare both RMSD and AAR. Finally ESM-2 to calculate Perplexity Pseudo Log Likelihood, proxy for biological relevancy.
The Job Sequencing And Tool Switching Problem
OPERATIONS RESEARCH
Sequence jobs so that you minimize the amount of tool switches you need to do. Solved using local search.
Introduced a novel local search operator that does ruin and recreate. It ruins and recreates by using a strategy that clusters jobs and best inserts local sequences. Aided by a well established constructive heuristic and guided using simulated annealing as meta heuristic.
