Embedded AI Project

NVIDIA Jetson Guide for Real-Time Deep Learning Inference from Scratch

Ioannis Dasoulas
A detailed guide to set up NVIDIA Jetson devices in order to run custom deep learning models from scratch efficiently. The guide contains instructions on booting embedding devices, configuring internet connection and memory, as well as setting up Python and running custom deep learning code.

Computer Vision Project

Mars Rock Segmentation and Size Estimation for Autonomous Rovers

Ioannis Dasoulas
Computer vision project that aims to identify and segment rocks in a Mars-like environment. YOLOv8 is used for rock detection and segmentation, as well as the MarsData dataset. Project is developed with pytorch, ultralytics, supervision and cv2 packages. For size estimation, a depth camera is used to estimate the size of the rocks. Model is also deployed on an NVIDIA Jetson Nano device, allowing real-time detection on rover cameras. Code will soon be open-sourced.

Data Engineering Project

Automated Feature Enrichment for Tabular Data with Semantic Table Annotation

Ioannis Dasoulas
Data engineering demo that leverages semantic table annotation to analyze tabular data, providing information about the column cells, types and relationships. The semantic annotations are used for querying external large knowledge bases and extracting more features about the table. In this demo we start with a table of airplanes, airplane families and airlines and manage to automatically extract new features such as wingspan, plane length, height and maximum operating altitude.

Natural Language Processing Project

Simple Trigram Language Model for City Name Generation with PyTorch

Ioannis Dasoulas
This project explores the usage of basic n-gram neural networks for character generation. A trigram neural network is trained on character pairs from greek city names to produce new city names. The project is an exercise task inspired by makemore's challenges. PyTorch is used for model development and training.

Robot Navigation Project

Point Cloud Surface Reconstruction for Autonomous Rovers

Ioannis Dasoulas
Robot navigation project that leverages point clouds inferred by rovers to reconstruct the surface and spaces around it. Different surface reconstruction and point cloud processing methods are tested. Point cloud segmentation are also leveraged to cluster the incoming point clouds. Open3D package is used.

Machine learning project (Diploma Thesis)

Chess Engine Development using Deep Learning Techniques

Ioannis Dasoulas
Three chess engines (Python Machine Learning Chess Engine) based on machine learning, and more specifically on reinforcement and supervised learning. The chess engines were developed using Python programming language. In total, three different game models were created, a model that was trained by playing games against itself, a model that was trained with data from past professional games stored in online databases and a combinational model that was trained using both methods. The engines’ choice of move is made using Monte Carlo Tree Search, which is a heuristic algorithm for certain types of decision-making processes.

Web development project (Diploma Thesis Side Project)

Chess Playing Website

Ioannis Dasoulas
Online application that allows the user to play against the chess engines mentioned on the above project, but also to watch them play against each other. The website was created using, Python (Flask API), HTML5, CSS and Javascript.

Database Application

Energy Distribution System Application

Ioannis Dasoulas, Aristeidis Dourdounas
Creation of an app about an energy distribution system. The app is meant to be used by a company that manages the distribution of electrical energy. The app provides a general picture of the distribution network, providing information about citizens' energy consumption and the energy produced by indepedent units of renewable energy. The project is written in Python and SQL.

Web development project

Smart-Home Application

Christoforos Vaggelis, Ioannis Dasoulas
Application meant to simulate a real smart home application which enables the user to control smart devices in his/her house and assign each device to a room of his/her house. The application provides three access levels for three different classes of users, assigned by the administrator of the house. Application is written in HTML5, CSS, Javascript, using Node JS and express.js for backend as well as Bootstrap for frontend. The website is uploaded on Azure websites (currently offline).

Machine learning project

Real Estate Value Prediction Algorithms

Ioannis Dasoulas
Python project using 9 house characteristics such as location, area population and number of rooms to predict the value of houses. The prediction is accomplished using: a) Least Mean Squares Linear Regression, b) Ordinary Least Squares Regression and c) Neural Network.

Human-Computer interaction project

MyDidaskalia Application

Ioannis Dasoulas, Christoforos Vaggelis
Zoom optimization project about designing an application named "MyDidaskalia" used for modern online lessons. The first part of the project was evaluating well known applications used for this purpose. The second part was choosing one of those applications (Zoom was chosen) and trying to improve it following human-computer interaction rules.

Data retrieval project

Movie Search Engine Application

Ioannis Dasoulas
Movie search engine based on elasticsearch. The search engine decides the order of results' presentation utilizing machine learning techniques. The order of presentation depends on elasticsearh similarity metrics, user ratings, average ratings and user clustering. Word embeddings were also used to predict missing ratings of certain movies for the users.

Parallel programming project

OpenMP Utilization in Classic Algorithms

Ioannis Dasoulas
Four projects about the utilization of OpenMP library in parallel programming. The library is used to accelerate the runtime of the classic K-Means algorithm, the travelling salesperson problem and the error-backpropagation computation in neural networks. All algorithms are written in C language.

Java multi-thread project

Sleeping Barber Problem Application

Ioannis Dasoulas
Classic sleeping barber problem implementated in Java language using object oriented techniques and thread syncronization.

Java server-client project

Blood Pressure Monitoring System Application

Ioannis Dasoulas
Application that has both server and client modes. Users can send medical data through the client interface to the server or use the server interface to view data sent by the clients, sort them, find spesific details, delete them or save them to a local file.

VLSI project

VHDL Circuits Simulation

Ioannis Dasoulas
Classic circuits design and simulation using VHDL language and the ModelSim environment.

Digital Signal Processing project

C67 DSK Board Programming

Ioannis Dasoulas, Aristeidis Dourdounas
C67 DSK board programming using Assembly and C language. The board was used for interrupts experimentation, A/D and D/A conversion, sampling, creation of FIR and IIR filters, creation of a DTFM decoder and creation of a spectrum analyzer.

Pattern Recognition project

Classic Pattern Recognition Algorithms Implementation

Ioannis Dasoulas
Pattern recognition projects using Matlab language. Projects include pattern sorting using distance functions, K-Means algorithm implementation, Perceptron algorithm implementation and data classification.

Data Mining project

Mail and Patient Classification

Ioannis Dasoulas
Mail classification (real or spam) and patient classification (possible stroke or not) using data mining and artificial intelligence techniques. Techniques include creating a K-Nearset-Neighbors model, creating a Random Forest Classifier model, as well as word embeddings with the implementation of a Doc2Vec model supported by a neural network.

VLSI project

Circuit Design and Simulation

Ioannis Dasoulas, Christoforos Vaggelis
Design and simulation of basic and complicated CMOS circuits, study of the parasitic capacities, the delay, the critical path delay and the energy consumption of CMOS circuits using the Microwind tool. Also, LtSpice was used to design some of the circuits and simulate them.