Aleksandar Gavrić

I am a PhD researcher at TU Wien, working at the Business Informatics Group under Prof. Dominik Bork and Prof. Henderik Proper.

My work focuses on world-model construction by integrating multimodal process mining, mixed-reality simulation, and video generation models to capture and model tacit expertise.

Previously, I worked with the Lydia Kavraki's, and Edward Knightly's Labs (Rice University) through internships, Mackenzie Mathis (EPFL) as a TA in the DeepLabCut course, Aleksandar Stanimirovic (Univesity of Nis) on bachelor and master thesis, and founded AI-for-Hotels.com, where I developed RAG, robotics and holographic systems.

Email  /  CV  /  Bio  /  Scholar  /  X.com  /  Github

Research

I'm interested in conceptual modeling, world-model construction, computer vision, deep learning, generative AI, multimodal representation learning, and differentiable algorithms for discovering, understanding and simulating real-world processes. Some papers are highlighted.

Built a structure-conditioned video generation pipeline using Petri nets as causal scaffolding for diffusion models.

A unified research framework integrating multimodal process mining with mixed-reality elicitation. The thesis models how real-world work unfolds by combining video, audio, interaction logs, sensor data, and immersive MR simulations to extract tacit expertise and build next-generation conceptual models.

Proposed embedding-based multimodal fusion and crossmodal alignment to enrich conceptual models.

Converts BPMN models into executable synthetic video animations for automated evaluation.

Proposes an alignment framework that tunrs multimodal embeddings into formal conceptual model symbols, enabling explainability, traceability, and hybrid reasoning in mixed human–AI modeling workflows.

Introduced a Petri net–guided prompting paradigm improving process-aware reasoning.

Defined process mining on latent AI representations; introduced relaxed discovery and conformance checking.

Introduces ViEnNa comics as a process-model notation that combines object-centric event logs and multimodal evidence into narrative visual diagrams enabling richer, more intuitive process understanding.

Real-time surgical multimodal process mining and MR-guided conformance checking.

Applied vision–language models to relate UML diagram elements with synthetic visual or acoustic evidence; introduced a user study linking UML fragments to verbal descriptions and observed interactions.

Designed NLP pipelines that translate multimodal observations into stakeholder-specific jargon, enabling domain-adaptive representation and improved explainability.

Augmented traditional event logs with multimodal signals; developed crossmodal encoders to infer missing or tacit activities.

Built physics-informed models predicting transistor degradation; combined simulation data with ML-supported curve-fitting for accuracy under thermal and stress conditions.

Formulated multimodal process mining as a representation-learning task; developed unified embeddings combining video, audio, and UI interactions for robust discovery under ambiguity.

Surveyed ML-ready encodings of BPMN/UML/Petri nets; categorized graph-neural, image-based, and text-based approaches; highlighted open challenges in multimodal interoperability.

Demonstrated XR-based reenactments for generating high-quality multimodal logs; evaluated video-driven activity recognition for process discovery.

Examined health issues associated with XR headset usage and introduced a system for detecting and monitoring ocular and vestibular complications, including actionable guidance for prevention.

Surveyed how adaptive optics, eye-tracking, and light-field display physics can enable immersive XR experiences without traditional corrective lenses.

Showed experimentally how spatial and temporal dimension reduction of sensor streams can lead to more successful predictive models in different applications.

Designed and implemented a predictive ML model for a distributed system applying machine learning on data streams to estimate dominant pollution sources in real time.

Miscellanea

Website template basend on ✩.