About me

I am a very friendly person, enjoying music, sports (anything involving a ball), board games (currently very much into Dune Imperium board game) and other social activities. I like travelling and exploring. I am an avid learner, especially about topics I never studied before.

Currently, I work as an AI research scientist at DeepL, focusing on improving our reinforcement learning efforts for aligning our LLMs closer to the customer's needs. Fun fact about the job is that I read many more papers nowadays than I had in the latter half of my PhD, hence the pace of learning is quite quick and the incorporation of the cutting-edge research into our codebase is done with utmost alacrity.

Before this, I was a postdoc at the Max Planck institute for astrophysics in Garching, Germany. Focus was on binding the developed code-base from the C++ environment into python (read JAX), in order to utilize the capabilities of numpyro library, and automation of chain convergence diagnostic also developed during my PhD.

I did my PhD at the Max Planck institute for astrophysics in Garching, Germany with Fabian Schmidt. Work was focused on the development of novel Bayesian forward modelling approaches for extracting useful cosmological information from the large-scale structure data, utilizing the effective field theory of structure formation. I am also a part of the Aquila consortium, and have contributed to several projects within the consortium. You can access my PhD thesis here.

My master thesis work was done at the same institute, however in the group of Torsten Ensslin, with additional mentoring from Reimar Leike. This was within the theoretical physics master program at the Ludwig Maximilians University in Munich. I focused on learning about the connections between differential geometry and learning algorithms, modelling causal relationships within bivariate datasets and quasi-periodic signal extraction using Gaussian processes with the NIFTy package. From this thesis, several publications arose as a result and several enhancements of the package itself. The thesis itself is available at these overleaf coordinates. The accompanying code can be found on my github repo.

I finished my Bachelor studies in Belgrade, Serbia at the Belgrade University, Faculty of mathematics. I focused on theoretical physics and astrophysics classes. Student's are not obliged to submit a bachelor thesis at the end, however I myself tried to do a small research project every summer. I listed all those projects below in the Unpublished projects section.

Publications

How much information can be extracted from galaxy clustering at the field level?
Nhat-Minh Nguyen, Fabian Schmidt, Beatriz Tucci, Martin Reinecke, Andrija Kostić
pdf| abstract

No evidence for p- or d-wave dark matter annihilation from local large-scale structure
Andrija Kostić, Deaglan J. Bartlett, Harry Desmond
pdf| abstract

Consistency tests of field level inference with the EFT likelihood
Andrija Kostić, Nhat-Minh Nguyen, Fabian Schmidt, Martin Reinecke
pdf| abstract

Constraints on dark matter annihilation and decay from the large-scale structure of the nearby universe
Deaglan Bartlett, Andrija Kostić, Harry Desmond, Jens Jasche, Guilhem Lavaux
pdf| abstract| blog post

Towards Moment-Constrained Causal Modeling
Matteo Guardini, Philipp Frank, Andrija Kostić, Torsten Ensslin
41st MaxEnt2022 | International Conference on Bayesian and Maximum Entropy methods in Science and Engineering
abstract

Causal, Bayesian, & Non-parametric Modeling of the SARS-CoV-2 Viral Load Distribution vs. Patient's Age
Matteo Guardini, Philipp Frank, Andrija Kostić, Gordian Edenhofer, Jakob Roth, Berit Uhlmann, Torsten Ensslin
pdf| abstract| code

Optimal machine-driven acquisition of future cosmological data
Andrija Kostić, Jens Jasche, Doogesh Kodi Ramanah, Guilhem Lavaux
pdf| abstract

Unpublished projects

Galaxy image modelling using shapelets with sparse techniques

with Arun Kannawadi. The goal was to use insights from the algebra of quantum mechanics operators and apply it to image processing of galaxies in order to determine their shape as acurately as possible. Determining the shapes comes especially handy for weak-lensing applications. Essentially, reducing down to Hermite polynomial decomposition of images with a clever basis construction. This was a project within the LEAPS 2017 program at the Leiden University, Netherlands.
pdf| code

Hunting for Intermediate-Mass-Black-Holes (IMBH)

with Paolo Bianchini and Glenn van de Ven. Goal of the project was to use internal dynamics, in particular the velocity dispersion, of globular clusters for finding hints of IMBH presence within the cluster. Work was done as part of a 2.5 month research internship at Max Planck institut for Astronomy in Heidelberg, 2016. Coding was done mostly inside Python/C/C++.
pdf

Examining the Kuiper belt sculpting

with Pedro Lacerda and Sebastian Lorek. Aim was to investigate the consequences of the Nice model towards the Kuiper belt structure. The work was done as a part of a month long research internship at Max Planck institut for Solar system research in Göttingen, 2015. Coding was done inside Fortran/C++ and bash. Some results can be found here.

Dynamics of Dust Particles Ejected from the Nuclei of C/2012 S1 (ISON) and C/2011 W3 (Lovejoy) Comets

with Mateja Bošković and Igor Smolić. This started as a high-school project within Petnic Science Center and involved implementation of internal thermodynamics model of comet nuclei as well as a small N-body integrator. Project was done almost exclusively within C with a flavour of MATLAB for visualising the results.
pdf| code

Belgrade University
2014 - 2018
MPS Göttingen
S2015
MPIA Heidelberg
S2016
Leiden University
S2017
Ludwig Maximilians University
2018-2020
MPA Garching
2020 - 2023
DeepL
2023 - present