Previous Projects
Master’s Thesis: Neural Representations of Faces in the Human Brain
My master’s thesis, supervised by Dr. Ben Harvey, investigated the neural mechanisms of face recognition in the human brain. The project leveraged the large-scale Natural Scenes Dataset (NSD), a comprehensive fMRI dataset, to overcome the limitations of traditional neuroimaging studies and provide a deeper understanding of how faces are represented in the visual cortex.
By applying Representational Similarity Analysis (RSA) and other advanced computational techniques, I developed a versatile analysis pipeline to identify and characterize face-selective brain regions. The research confirmed that early face representations are primarily based on spatial configuration, similar to findings in non-human primates. I also compared these findings with neural network models to identify fundamental differences between biological and artificial systems in how they process and represent faces.
This project demonstrates my expertise in fMRI data analysis, computational neuroscience, and machine learning. I utilized a suite of technologies including Python and neural networks to produce novel insights into the hierarchical organization of object recognition in the brain.
Integrating AI into business processes
During my role at MicroNova, I was responsible for developing and integrating two distinct AI-powered applications to streamline key business operations. Both projects involved building end-to-end solutions, from data processing to user-facing applications, demonstrating my ability to deliver production-ready systems that solve real-world problems.
Automated Contract Analyzer
The Automated Contract Analyzer is a production-ready application that classifies and extracts key information from contracts. I developed a custom Retrieval-Augmented Generation (RAG) application using LlamaIndex and ChromaDB to pre-fill contract details and provide a conversational interface. This system significantly reduced manual data entry and improved accuracy by providing direct references from the source documents to build trust in the extracted information.
Key Responsibilities & Technologies:
-
Designed and built the full RAG pipeline from scratch.
-
Developed a document processing pipeline to handle various file types (PDF, CSV, text) and extract data using Optical Character Recognition (OCR).
-
Optimized query parameters (temperature, top_k) and integrated a reranker for improved accuracy.
-
Provided an API for seamless integration with existing business systems.
-
Planned for cloud-based scaling on Azure.
Audit Helper
The Audit Helper is a Microsoft Word Add-in designed to assist with audit processes. I created a solution that transcribes recorded audit conversations and makes them searchable and manageable via a user-friendly interface. The system uses a Speech-to-Text model and a RAG-based conversation engine to help auditors quickly find and insert relevant information into their Word documents.
Key Responsibilities & Technologies:
-
Developed a Word Add-in using HTML, CSS, JavaScript, and a Flask backend.
-
Implemented a system for transcribing large audio files with Whisper, optimizing the process with FFmpeg and Opus Codec.
-
Integrated speaker diarization to distinguish different speakers in the conversation.
-
Built the front-end to display transcription and RAG-based conversations, with automated functionality to insert content directly into the Word document.
Computer Vision & Remote Sensing at Fraunhofer IGD
During my time at the Fraunhofer Institute for Computer Graphics Research (IGD), I gained extensive hands-on experience in computer vision and deep learning through various roles, including an internship, research assistantship, and my thesis project. My work focused on developing and evaluating state-of-the-art models for a variety of tasks related to remote sensing.
My responsibilities spanned the entire machine learning lifecycle, from data processing to model training and evaluation. I trained both Convolutional Neural Network (CNN) and Vision Transformer (ViT) models using PyTorch/torchvision and the OpenMMLab framework to tackle challenges such as image-based regression, keypoint detection, and object detection.
I also explored advanced research topics, including Multi-Task Learning, Data Fusion, and AI Explainability, specifically by analyzing attention maps from ViT models. To ensure a structured and reproducible workflow, I utilized MLOps tools like MLFlow for experiment tracking and model management.
This comprehensive experience has provided me with a deep understanding of computer vision, proper software engineering practices, and the end-to-end development of deep learning applications. It demonstrates my ability to contribute to complex research and development projects.
Evolutionary algorithms: Improving traditional heuristic algorithms
This project tackles the Bipartite Graph Partitioning Problem by comparing two approaches:
- Fiduccia–Mattheyses (FM) heuristic: a fast local search that iteratively moves vertices to reduce cut size.
- Genetic Local Search (GLS): an evolutionary algorithm that combines crossover/mutation with FM-based local refinement to escape local optima.
By extending a classic local search heuristic with evolutionary concepts, this project shows how population-based search and local refinement can complement each other.
The evolutionary approach consistently improves solution quality, avoids local optima, and demonstrates that even well-established heuristics can be pushed further by combining them with meta-heuristic strategies.