Whether it comes to management, data science or programming, I am always seeking innovative approaches to solve problems.
You can find a list of personal projects below, showcasing some of the technologies that I am passionate about, together with the links to more detailed documentation.
For a clear and concise representation of my past experiences, you can also check out my Linkedin.
If you still have some questions, I will be happy to have a chat!
The project is a Javascript adaptation of a processing application that I previously wrote using Python, Jinja and Flask. This serverless implementation was a good playground to experient with frontend technologies and to introduce new algorithms into the workflow.
The editor features a Quantizer, taking care of the reduction of the colors of the uploaded image to a user-defined palette. A Flood Fill recursive algorithm allows to select different brush sizes in the editor and a "paint bucket" function.
Finally, to overcome the limitation of intermittent registration of the cursor location while in drawing mode, a custom implementaiton of Dijkstra's algorithm was used. This allowed to find the distance travelled by the cursor between subsequent mouse location events and draw a continuous path between them.
Next steps: I am currently working on the integration of a cGAN (Conditional Generative Adversarial Network) within the quantization algorithm, to dynamically modify the original picture to look like retro pixel art.
Fun fact: The cover image of this modal, "LEGO" effect included, was created with the Pixpad editor.
Algorithms, Visualization, Frontend Architecture
This demo was develped using Tensorflow.js, Firebase and React as an experiment to integrate a deep learning model and realtime database access into a frontend application.
To ensure an optimal experience, please use the desktop version of the Google Chrome browser.
This implementation is based on a training instance of the COCO dataset with over 80 labels (e.g. vehicles, animals and household objects). The experiment introduces several improvements over my previous computer vision project for real-time face recognition, written in Python and based on a Multi-task Cascaded Convolutional Neural Network (MTCNN).
Algorithms, Visualization, Frontend Architecture
Using the connected webcam the captured frames are processed by a Multi-task Cascaded Convolutional Network (MTCNN) that predicts the face location and Proposal Network (P-Net) is used to obtain the facial windows and their bounding box. All detected faces are fed to a different CNN, called Refine Network (R-Net), which reduces the number of false positives. After the preparation steps are completed, the framework uses Facenet to extract features to be used in the classification. The framework can detect also multiple people in a single frame.
Deep Learning, Machine Learning, Python
The analysis is provided in a Jupiter Notebook format and was first submitted for a Kaggle competition on the public heart desease prediction dataset from the Hungarian Institute of Cardiology.
Machine Learning, Data Visualization, Feature Selection, Model Optimization
For this project I created an Instagram bot to access a profile, gather information regarding followers and engagement and automatically like and follow people with similar interests and hashtags. The bot collects and stores information regarding liked pictures and followed profiles allowing to maximize the followed-to-followers ratio. Followed people that do not increase the engagement beyond a certain treshold are automatically unfollowed after a set time.
Web scraping, Analytics, Selenium, Robotic Process Automation
The goal of the project was to automate the manipulation, classification and sorting of wholesale invoices. The featured script is part of an internal production application developed for the finance back office of one of the "tech big four". The present script is complemented by an NLP classifier extracting the key information for different invoicing and rebate documents. All the processed documents, after being classified, are saved to the correct location in a local database and processing logs are generated for debugging and analytics purposes.
OSX Scripting, Natural Language Processing, Integrated Applications
