Hi! I’m Josip, I work in Zalando, a Berlin-based fashion and lifestyle e-commerce company, and Europe’s biggest fashion online retailer. I work on hybrid and neural information retrieval, interactive search and information retrieval evaluation in e-commerce using multi-modal large language models.

I am also interested to learn about biology and material science, especially about applications of generative models e.g. autoregressive generative models for design of sequences that define objects (e.g. ligands, proteins, RNAs, materials) that behave in a desired way, e g. perform a certain function (e.g. bind to proteins) or have a certain property (e.g. conductivity). The objects’ behavior is determined by its structure (e.g. 3D shape of folded proteins or crystal lattice structure of materials) and the structure is defined by the chemical composition and chemical and physical laws.

This composition-structure-function/properties mapping can be modeled by a model, e.g. a deep neural net. These models can be trained from data, taking into account constraints defined by chemical and physical laws and are made data efficient by including symmetries of the 3D space. The inner states of these models correspond to representations of objects described by the input sequences. Once trained they can not only help us predict the structure and properties of objects defined by unknown sequences, but also find objects with similar properties and structure and generate sequences that describe objects, given the desired structure or properties as inputs.