Anastasios Stamoulakatos

Anastasios Stamoulakatos

Data Scientist at Satalia | PhD in Applied AI

Applied Research

WPP Open ResearchLab: Data Quality Assurance Agent

Published:

  • Research and development of an intelligent AI agent for automated data quality assurance and schema validation.
  • Implementing reasoning-based agents to analyze databases, identify anomalies, and enforce consistency constraints.
  • Part of the WPP Open ResearchLab initiative.

Panoptic Segmentation with Mask2Former & PyTorch Lightning

Published:

  • Training, evaluating, and running inference for Panoptic Segmentation models using Mask2Former.
  • Leveraging PyTorch Lightning and Hugging Face Transformers for modular and scalable training pipelines.
  • Built during Track32 AI research and development.

PyTorch 3D IBN-Net

Published:

  • 3D Convolutional Neural Network (3D CNN) incorporating Instance-Batch Normalization (IBN-Net) blocks.
  • Designed for video spatiotemporal analysis and action recognition with enhanced domain generalization.
  • Developed during PhD studies to improve feature extraction robustness under varying subsea conditions.

Custom Cutout Data Augmentation in PyTorch

Published:

  • Implementation of Cutout data regularization/augmentation for convolutional neural networks in PyTorch.
  • Provides modular and flexible cutout masks to improve model generalization and robustness against occlusion.
  • Developed during PhD studies to enhance data variety for pipeline inspection datasets.

PyTorch Video Dataset Loader & Sampler

Published:

  • Efficient dataset loader and sampler helper classes for video classification tasks in PyTorch.
  • Optimizes random sampling, clip loading, and preprocessing routines for heavy video data pipelines.
  • Used extensively during PhD projects to accelerate model training on subsea survey footage.

PyTorch Swapping Autoencoder

Published:

  • Implementation of Swapping Autoencoders for deep generative representation learning.
  • Enables texture and structure manipulation for image-to-image translation tasks.
  • Part of PhD research on unsupervised representation learning.

PyTorch IBN-Encoder LSTM-Decoder

Published:

  • Encoder-Decoder network architecture combining IBN-Net feature extraction with LSTM sequential modeling.
  • Designed for video frame-level prediction, subsea survey annotations, and sequential data analysis.
  • Developed during PhD studies to automate pipeline anomaly detection over continuous video frames.