Gergely Papp
Machine Learning Engineer
Expert in deep learning, computer vision and NLP.
Expert in deep learning, computer vision and NLP.
Python
SQL
JavaScript
Java
C#
PyTorch
TensorFlow
Keras
Huggingface
Scikit-Learn
Pandas
NumPy
OpenCV
Django
Flask
NodeJS
Docker
Git
Cloud
Azure
Databricks
REST API
SLURM
Neptune.ai
Wandb
Jenkins
CI/CD
I confidently execute end-to-end ML projects across computer vision and NLP domains, with publications at top conferences backing my expertise. While grounded in fundamental ML theory and practice, I'm actively expanding my knowledge of industry-standard tools and platforms to enhance implementation efficiency and deployment capabilities.
2020 - 2024
2024
2020 - 2024
2018 - 2020
2017 - 2018
Grade: 8.1, Cum Laude
Grade: Second-Upper Class
Extracting psychological data, resilience, and player communication statistics from e-sport footage for further analysis.
Developing a natural language interface for terminal operations that can execute complex tasks through prompting, with optimization for deployment on resource-constrained devices.
Participated in a Retrieval Augmented Generation project involving vector databases, knowledge graphs, and text generation with LLMs.
Investigated Vision Transformers' ability to generalize across object properties (shape, texture, color, count) on CLEVR-4.
Analyzed how NLP transformers fine-tuned for different languages and tasks relate geometrically and functionally, revealing potential for cross-task insights.
Developed a rule-based AI in NodeJS for automating gameplay for a browser game with attack timing, reaction to reports, and HTML dashboard logging.
Developed a time series forecasting model to predict energy consumption based on historical data through collection, cleaning, preprocessing, and model training.
Explored innovative self-supervised image classification methods competing with state-of-the-art approaches, achieving promising results on smaller datasets.
Demonstrated that geometric and functional similarity in neural networks are distinct concepts using affine transformations between networks.
First NeurIPS paper accepted from a Hungarian institute (2020), establishing a novel approach to comparing neural network representations beyond geometric similarity.
Visualized the role of neurons in CNNs using Lucid and GANs, revealing what images best represent specific classes in a CelebA-trained classifier.
Initiated and built a ViT-based license plate recognition system during a brief window of opportunity while management was away.
Successfully replaced an expensive third-party solution with a superior in-house system that became a core product for Asura Technologies, demonstrating initiative and technical excellence.
Developed a system to extract Machine Readable Zones from passport images through comprehensive data collection, preprocessing, and model training.
Created an ML system to track quarantine compliance during the COVID-19 pandemic, with face detection and anti-cheating mechanisms.
Successfully delivered a critical application under tight time constraints that was adopted by the Hungarian government and used by tens of thousands of citizens.
Developed a real-time system to count wheels through comprehensive data collection, preprocessing, and model training.
Implemented a real-time YOLO-based system for detecting vehicles and license plates, overcoming challenges with fish-eye camera distortion through specialized preprocessing.
Developed a tool to count people entering a shopping mall in real-time with 90%+ accuracy, providing reliable customer traffic estimates.
Categorized car images into their make and model with 90%+ accuracy through comprehensive data collection, preprocessing, and model training.
Built an end-to-end OCR solution that automatically cleans, rotates, and extracts readings from watermeter images, deployed to the cloud for seamless integration.
Reimplemented AlphaZero to explore temporal difference learning vs. Monte Carlo methods. The study revealed unique in-game strategies made with Reinforcement Learning.
Engineered a tool using hierarchical clustering to reconcile balance sheets between different accounting systems, automating complex financial comparisons.
Architected an OCR system that translates invoices into structured formats and extracts relevant information using a shallow neural network for efficient data processing.
Investigated latent space properties of structured VAEs by forcing a torus shape while reconstructing clock images, advancing understanding of controlled generative models.
Tackled the challenge of detecting small firearms in high-resolution images in real-time, discovering the critical importance of contextual information for accurate detection.
Built a predictive model using linear regression to forecast future values based on historical time series data, enabling data-driven decision making.
Designed a Java-based neural chess engine from scratch without the use of tree search, achieving entry-level play
"Mode combinability: Exploring convex combinations of permutation aligned models"
"Reproducibility study of 'Label-Free Explainability for Unsupervised Models'"
"Similarity and Matching of Neural Network Representations"
"Ordering Subgoals in a Backward Chaining Prover"
4 times a week
8 years experience
When time permits
Every 2 weeks
Champion in 2010