Mathematics · Machine learning · Munich
I'm a mathematician who builds things. Master's in Mathematics in Data Science at TUM, bachelor's in Computer Science at KAIST, and before that, olympiad mathematics for Azerbaijan.
The dots behind this page run a diffusion model, the kind my thesis is about: noise in, shapes out, and back. Your cursor adds noise; click (or tap) for a burst, or draw your own shape. Or play with it in the playground.
Technical University of MunichM.Sc. Mathematics in Data Science · Munich, Germany
Master's thesis on discrete diffusion language models at the Data Analytics and Machine Learning group (below).
KAISTB.Sc. Advanced Major in Computer Science, Minor in Mathematics · Daejeon, South Korea
Korea Advanced Institute of Science and Technology, on a full merit scholarship. Computer science with a minor in mathematics and a semi-minor in AI.
DAML, Technical University of MunichMaster's thesis · Munich
Discrete diffusion language models. Controlled ablations showed that DUO's temperature curriculum is inert and its argmax trajectory invisible to the training objective, tracing the reported speed-up to an induced noise schedule. Principled noise samplers then reached the target perplexity 2× faster than DUO, with 139M-parameter models trained on distributed GPUs.
PyTorch · Hydra · Slurm
IAVData scientist, working student · Berlin
Reliability statistics for vehicles: a generator of synthetic failure-lifetime data from Weibull mixtures, benchmarks of survival-analysis methods (log-rank, Kolmogorov–Smirnov, Kaplan–Meier), and an analytical model, derived by hand, that replaced a slower Monte Carlo simulation. Also compared Gaussian-process frameworks (GPyTorch, GPy, GPflow, KeOps).
Python · Docker · CI/CD · Kubernetes
ZEISS × TUM DI-LABML researcher, part-time · Munich
Continual learning for medical image segmentation with Mask2Former. Benchmarked replay, knowledge distillation and contrastive learning against catastrophic forgetting: pixel-wise supervised contrastive learning with replay reduced it substantially. Designed a weighted contrastive loss and prototype anchors for severe class imbalance.
PyTorch · Hugging Face · project page ↗
Saesol TechFull-stack software developer · Seoul
Cybersecurity for connected vehicles (V2X): a test station for the certificate system SCMS, presented at Florida PlugFest in May 2023; the bootstrapping that gets a vehicle device its certificates; and butterfly key expansion, which turns one certificate request into many certificates.
C · C++ · FLTK · gRPC
BionSightBioinformatics engineer, intern · Seoul
Re-implemented the molecular docking program AutoDock Vina (C++) in Python, with a Monte Carlo iterated search and a BFGS optimiser.
Python · NumPy · SciPy · RDKit
Software Security Lab, KAISTResearch intern · Daejeon
Binary lifting: extracting LLVM bitcode from compiled binaries with B2R2 (F#), including a disassembler for the AVR architecture written from scratch.
For Azerbaijan, 2015 – 2018: six medals at the International Mathematical Olympiad, the Balkan and Junior Balkan Mathematical Olympiads and the International Zhautykov Olympiad. Cards with ↗ open the official results.
Three decades of UFC fights as explorable models: a 3D style map of every active fighter, a fight predictor, an Elo ladder, and Pick'em, a live prediction game with leagues and the model as a player.
Python · scikit-learn · React · Supabase · source ↗
Write to anar.abbasli@tum.de, or find me on GitHub and LinkedIn.
