MSc Computational Neuroscience at BCCN Berlin. I bridge machine learning and brain science — explore the intersection of neuroscience, machine learning, and cognitive science. Building models that predict human perception and higher brain functions.
I am a researcher at the intersection of computational neuroscience, computer vision, and cognitive science, currently completing my MSc at the Bernstein Center for Computational Neuroscience (BCCN), Berlin. My primary research interest concerns the computational principles underlying human visual attention — specifically, how structured semantic representations of dynamic environments relate to the goal-directed nature of gaze behaviour.
As a Student Assistant at the Science of Intelligence Excellence Cluster (SCIoI, TU Berlin), I develop end-to-end pipelines that link open-vocabulary scene graph representations with human gaze scanpaths. This work combines vision-language models for relational scene parsing, GloVe-based predicate embedding, and K-medoids clustering to investigate which semantic relation families in dynamic scenes are preferentially selected by human visual attention. During my lab rotation at the NeuroImaging Group (BCCN), I applied multivariate pattern analysis (MVPA) to real fMRI data, evaluating experimental design validity through the Same Analysis Approach (SAA) and The Decoding Toolbox (TDT) within the SPM framework.
Prior to my transition into neuroscience research, I accumulated over a decade of experience in software engineering and systems architecture, with a focus on large-scale SaaS development and enterprise-grade distributed systems. This background informs a rigorous, reproducibility-oriented approach to computational research.
My broader interests lie in the representational basis of perception and cognition — how symbolic, geometric, and neural representations can be integrated to model the selectivity and flexibility of biological intelligence.
Building a pipeline that links open-vocabulary scene graphs from dynamic videos with human gaze scanpaths (Project 57). Constructs frame-wise scene graphs via VLM, extracts foveation-based temporal relations from eye-tracking data, embeds predicates using GloVe, clusters them with K-medoids, and identifies which semantic relation families are preferentially selected by human visual attention.
Verified · SCIoI Profile PageAnalyzed fMRI data using the Same Analysis Approach (SAA) and The Decoding Toolbox (TDT) within SPM. Created a tutorial demonstrating these tools for the lab.
Led agile delivery across 4+ teams (~30 people) in an Enterprise Service Bus project, collaborating with Product Owners to align technical delivery with customer needs.
Developed SaaS banking solutions (Card Management System). Specialized in synchronized programming, data access control, and open banking APIs.
Contributed to the Central Deposit System using REST/SOAP architecture in a cross-functional team.
Exploration of graph-based conceptual representations — building structured semantic graphs connecting concepts and their relationships. Part of ongoing interest in combining symbolic and neural approaches to knowledge representation.
Research question: Which semantic relation clusters in an open-vocabulary scene graph are preferentially selected by human gaze in dynamic environments?
This project investigates the link between structured scene representations and human visual attention. The full pipeline constructs open-vocabulary scene graphs from dynamic videos, extracts gaze-based temporal relations from eye-tracking scanpaths, and compares which semantic relation types humans preferentially attend to — distinguishing between relations that are merely frequent in the environment versus those actively selected by attention.
Goal: Evaluate the validity and decoding reliability of an ongoing lab experiment using real fMRI data.
As part of my lab rotation at the NeuroImaging Group (BCCN Berlin), I worked with real fMRI data from an active research project. I applied the Same Analysis Approach (SAA) to assess whether the experimental design could support robust, reproducible decoding — a principled method that tests whether the same analytical pipeline succeeds on real data but not on permuted controls, directly probing design validity before drawing scientific conclusions.
Decoding was carried out using The Decoding Toolbox (TDT) within the SPM framework, applying multivariate pattern analysis (MVPA) to identify brain regions whose activation patterns reliably distinguish experimental conditions. As a final deliverable, I produced a step-by-step tutorial covering the full SAA–TDT workflow to help the lab run reproducible analyses in future projects.
Problem: Predicting bank customer satisfaction from a high-dimensional, severely class-imbalanced dataset (96% satisfied vs. 4% unsatisfied) using an interpretable rule-based learning system.
Applied a Michigan-style Learning Classifier System (LCS) — a genetic-based machine learning approach that evolves a population of IF:THEN rules — to classify Santander Bank customers. The raw dataset contained 369 anonymized features across 147,392 records. A rigorous preprocessing pipeline handled class imbalance via SMOTE, removed constant and highly-correlated features (threshold: 0.85), and converted all attributes to binary format for rule-based processing.
Three rule compaction strategies were benchmarked — QRC, QRF, and PDRC — using 5-fold cross-validation. QRC achieved the highest accuracy (88.4%), while QRF was the fastest. Attribute Tracking and Feedback mechanisms further identified the four most predictive features driving customer dissatisfaction.