Schlagwörter
Skills
I have worked on projects for various clients in different industries, using my expertise to help the organisation improve efficiency, reduce costs, and increase revenue through the use of data-driven solutions.
Frameworks:
- Keras, PyTorch, scikit-learn, TensorFlow, XGBoost
- Conda/Anaconda, Jupyter, Matplotlib, NumPy, openCV, pandas, plotly, Poetry
- MLflow, SageMaker, Vertex AI
- Anomaly Detection, Audio Analysis and Synthesis, Clickstream Analysis, Computer Vision, Content Understanding, Data Analysis, Data Mining, Data Visualisation, Deep Learning, Dynamic Pricing, Fraud Detection, Image Processing, Image Recognition/Classification, Machine Learning, Natural Language Processing (NLP), Natural Language Understanding, Product Similarities, Recommendation Systems, Speech Recognition
- Deep Neural Networks, Convolutional Neural Networks, LSTM, (Variational-)Autoencoder, Transformers
- Hyperparamer Tuning, Transfer Learning
- Model/Feature Analysis using SHAP
- Dimensionality-Reduction (PCA, t-SNE, LDA, Autoencoder, UMAP)
- Python
- C/C++, Java, MATLAB/GNU Octave, PHP
- Clean Code, PyTest, Static Code Analysis, Unittest
- Bamboo, Bitbucket, Jenkins, Git, GitHub, GitLab
- Software Development and Software Architecture
- Linux, macOS, Windows
- Apache Spark, BigQuery, Elasticsearch, Exasol, Graylog, Kibana, MS-SQL, MySQL, Oracle DB
- Amazon Web Services (AWS), EMR, SageMaker, Apache Spark
- Google Cloud Platform (GCP), BigTable, BigQuery, Vertex AI
- Hadoop, PySpark
- FFmpeg for Video Processing
- Docker
- Kubernetes
- Confluence, Jira, Miro, Slack, Teams, Trello
Projekthistorie
As a freelance consultant and expert in Machine Learning applications for "content understanding" I support the RTL Data Team in building the next generation multi-purpose platform “RTL+” in cooperation with Deezer using visual (video), audio and textual data.
The primary objective of this project is to derive additional metadata from the raw content, which can subsequently be utilized by downstream applications such as search, recommendation and personalization. The key challenge lies in establishing a reliable, scalable and production-ready state-of-the-art solution for a vast number of building blocks and crafting an execution pipeline on top of it.
Video based models: Aesthetic Ranking, Dominant Color Extraction, End Credits Detection, Face Detection, Image Quality Detection, Logo Detection, Mood Detection, Object detection and Recognition, Place Prediction, Scene and Shot-Boundary Detection, Shot Type Detection by using and optimizing both pre-trained and self-trained models.
Audio based models and solutions: Speech-to-Text transcriptions using Google’s Speech-to-Text API and Whisper from Open-AI on Podcasts and other audio sources and music identification.
NLP solutions: language detection (fastText), festivity detection, kids content detection, adult content detection, topic modeling (BERTopic), keyword extraction (KeyBERT) and text summarization.
Toolkit: Argo Workflows, Confluence, Docker, Elasticsearch, FFmpeg, GitLab CI/CD, Google BigQuery, Google Cloud Platform (GCP), Google Data Studio, Grafana, Hugging Face models, Jira, Jupyter/JupyterLab, Kafka, Kibana, Kubernetes, MLflow, NumPy, pandas, Poetry, Pub/Sub, Python, PyTorch, Scrum, spaCy, SQL, Streamlit, TensorFlow, Terraform
As a freelance consultant and expert in Machine Learning, Data Science and Deep Learning, I specialized in fraud recognition, product recommendation systems, image recognition/classification, anomaly detection, time series analysis and NLP. I guided agile projects from conception to production and maintenance & optimization.
I focused on eCommerce solutions that leveraged consumer data, product master data & descriptions, product images and sales transactions.
Product Similarity: The goal of the product similarity solution was to improve downstream system performance by identifying similar or related products for a given product, which could then be used as a benchmark or replacement product. The similarity was determined using various modalities, including visual similarity (via an image autoencoder), consumer behavior (using clickstream data) and product descriptions (by using NLP transformer-based models).
Toolkit: AWS, Bitbucket, Confluence, Jenkins, Jira, Jupyter, PySpark, Python, Scrum, TensorFlow
Dynamic Pricing: The goal of this project was to identify poor-performing products in an early stage, uncover any potential product issues, and determine the right actions (such as an optimal price change) to boost performance. The ultimate goal was to gradually replace the existing solution.
Toolkit: AWS, Bitbucket, Confluence, Jenkins, Jira, Jupyter, Matplotlib, PySpark, Python, Scrum, TensorFlow, XGBoost
Consumer Lifetime Value (CLTV): I was responsible for the conception, implementation, and maintenance of the historical and future monetary value attributed to individual consumers. This included regular extensions and adaptations (e.g. for new markets/brands) and deep dive analyses into the model's most important features.
The models, which were based on consumer behavior data, ran in production and were updated on a weekly basis for all consumers. The results (KPIs) were intensively used in downstream systems and for marketing campaigns.
Toolkit: Bitbucket, Confluence, Exasol, Jenkins, Jira, Jupyter, Matplotlib, Python, SHAP, Scrum, XGBoost
Visual Product Embeddings: I was responsible for the conception and implementation of a variational autoencoder based on product images. The source images were filtered, downscaled, and prepared for a convolutional neural network (VAE) that generated embeddings capable of capturing design elements of a product image. These embeddings were used to find similar products and also fed into downstream models to improve product-based models. The solution ran in production and was updated with new images on a weekly basis.
Toolkit: Bitbucket, Confluence, Exasol, Jenkins, Jira, Jupyter, Matplotlib, Python, SHAP, SQL, Scrum, XGBoost
Purchase Propensity Scores: I was responsible for the conception, implementation, and maintenance of a model for predicting consumer purchase intentions. The solution had been running very stably in production for a few years already and the results had made a significant contribution to marketing channels.
Toolkit: Bitbucket, Confluence, Exasol, Jira, Matplotlib, Python, SHAP, SQL, XGBoost
Kaggle Challenge: I participated in the "Histopathology Cancer Detection" Kaggle competition, where the goal was to identify metastatic cancer in medical images. My role was to bring state-of-the-art computer vision techniques to the team and to implement an ensemble of models for the submission. We finished in 26th place out of 1,149 competitors by using advanced (high-speed) training techniques and heavy image augmentations.
Technologies: Python, Jupyter, PyTorch, plot.ly, GitHub
Date: 2017
Technology: TensorFlow, Keras, Convolutional Neural Networks
Use case: Image detection, classification and metadata extraction for product images
Goal: Enrich metadata for product descriptions from images, finding outliers - helping content management teams to improve data quality
Date: 2017
Technology: TensorFlow, Keras, Convolutional Neural Networks
Use case: Prototype for detecting product variants on an image
Goal: Reduce manual effort, fully automate and scale processes
Date: 2016 Technology: TensorFlow, LSTM
Use case: Audio signal analysis and synthesis using Deep Learning
Goal: Various experiments to deconstruct and construct audio signals
Date: 2016
Technology: TensorFlow, LSTM, Convolutional Neural Networks
Use case: Classification and anomaly detection using Deep Learning
Goal: Labeling transactions, find anomalies, reduce manual work
Date: 2015
Technology: Random Forest
Use case: Product recommendation using Random Forest algorithm
Goal: Product recommendation optimized on long time revenue
Date: 2014
Technology: Regression
Use case: Fraud detection
Goal: Find similarities between new customer registrations to prevent multiple registrations (fraud/misusage)
Date: 2014
Technology: Apache Mahout
Use case: Product recommendation using Item-Based Collaborative Filtering
Goal: Recommend similar Products for known users based on user behavior on a marketplace platform
Date: 2014
Technology: Apache Mahout
Use case: Product recommendation using k-nearest neighbors algorithm
Goal: Show similar Products for unknown users on a marketplace platform
Zertifikate
Reisebereitschaft
Sonstige Angaben
https://www.xing.com/profile/Andras_Molnar
https://www.kaggle.com/andrasmolnar