Piotr Durniat

thyroID - AI System for Thyroid Tumor Malignancy Detection (GitHub TBA)

University Scientific & Implementation Project (3 semesters) | Team of 4

Python | PyTorch | PyTorch Lightning | OpenCV | Pandas | NumPy | DVC | Weights & Biases (WanBD)

• Developing a multimodal AI system for thyroid tumor malignancy assessment using ultrasound images and patient blood samples.
• Collaborating with an endocrinologist to build a custom-labeled dataset of ultrasound images, annotated with real post-surgical malignancy outcomes.
• Implementing a self-supervised learning model using Vision Transformers (DINO) with a student-teacher architecture for feature extraction.
• Developing a supervised multiple-instance learning (MIL) model for final malignancy classification, using extracted features.

Dependency Attack VEctor Finder (DAVE) – GitHub

Complex Networks Course Project | Team of 3

Python | PyTorch | NetworkX | NumPy

• Developed a system for detecting potential software attack vectors by analyzing package dependency graphs, inspired by the XZ Utils backdoor incident.
• Built a web scraper to collect and process package metadata and usage statistics.
• Designed and implemented dependency graph analysis to assess the impact of vulnerabilities based on package dependencies and contributor statistics.

Curiosity Mastcam Novelty Detection – GitHub

Probabilistic Graphical Models Course Project | Team of 3

Python | Pyro | PyTorch | NumPy | Scikit-Learn

• Implemented and compared three generative models (BiGAN, VAE, and Normalizing Flow) for novelty detection in multispectral images of Mars captured by the Curiosity rover.
• Trained models on the Mastcam dataset to learn feature representations and identify anomalous patterns in Martian terrain.
• Developed anomaly detection techniques by analyzing latent representations and evaluated model performance in terms of accuracy and performance.

Bike Detection and Tracking – GitHub

University Project in Collaboration with NeuroSoft | Team of 3

Python | PyTorch | TensorFlow Lite

• Developed an embedded system for bike detection, tracking, and counting on Raspberry Pi with Google Coral TPU and an onboard camera.
• Integrated a pre-trained CNN-based object detection model with DeepSORT tracking and a custom line-crossing algorithm for accurate bicycle counting.
• Researched TensorFlow Lite model to balance accuracy vs inference speed for real-time use.
• Conducted end-to-end development, including real-world data collection, processing, and evaluation.

Dollar Cost Averaging Simulation – GitHub

Bachelor’s Thesis Project

Python | TypeScript | Docker | MongoDB

• Developed a financial simulation application for the Dollar Cost Averaging (DCA) strategy, using real-world financial data for investment analysis.
• Designed the system architecture, covering frontend, backend, and database, focusing on scalability.
• Implemented DCA algorithm and developed an interactive user interface for real-time result visualization.
• Conducted comprehensive testing, deployment, and documentation.

Crosswalk Detection – GitHub

Python | PyTorch | YOLOv5 | DeepSORT | Google Colab

• Created a custom training dataset for a CNN using self-captured and internet-sourced images.
• Processed, labeled, and preprocessed training data to optimize model input and performance.
• Fine-tuned the model using Google Colab with a rented GPU.
• Evaluated system accuracy in real-world videos.
• Implemented YOLOv5 for object detection and DeepSORT for tracking.

Other Computational Algorithms & Parallel Computing Projects

• Parallel Computing on Beowulf Cluster – GitHub – Set up a Beowulf Cluster using virtual Linux machines; implemented a parallel Sieve of Eratosthenes with MPICH in Python, comparing performance with a serial approach.
• Multi-threaded Computing (Pac-Man Game) – GitHub – Developed a multi-threaded program with mutex-based synchronization; created a graphical interface in Python.
• Optimization Algorithms (TSP Solver) – GitHub – Implemented and evaluated multiple Travelling Salesman Problem (TSP) algorithms (Brute Force, Dynamic Programming, Branch & Bound, Simulated Annealing, Genetic Algorithm) in C++.
• Ray Tracing Implementation – GitHub – Developed a recursive ray tracer incorporating Phong shading, reflections, transparency, and refraction in Python.