May 2024 - Aug 2024

I was hired as a Research Assistant to develop a Gaussian Process (GP) ML model to estimate the age of S/Sq/Q-type asteroid surfaces. The GP model is trained with hyperspectral reflectance data that is acquired from various laboratory tests. This algorithm has potential to be used to analyze hyperspectral asteroid images from future space missions.

GP model was selected for this task as it has great potential even with a relatively sparse training dataset like ours. I conducted a comprehensive study in order to select the most optimal model structure, which eventually consisted of a Convolutional Neural Network (CNN) based Feature Extractor, and Hadamard Multitask Regression which are built on top of the base GP model. The GP model uses Constant Mean and Matern Kernel as its base components.

The Hadamard Multitask functionality was useful in our case because it can be used to train a separate initial model for each of the different asteroid types, and to combine those models into one single algorithm, which is optimized to benefit from the possible similarities between the asteroid types.

I'm proud of the results this model was able to achieve. The results were comparable with an ensemble model that uses CNN, Gradient-Boosting regression, K-Nearest neighbor, Extra-Tree Regression, and Random Forest algorithms. More information can be seen in the future once our report paper is published. In the mean time, there is a link to my Github repository below.

I learned:

• Mathematical Understanding of the Underlying GP and CNN Models
• PyTorch/GPyTorch
• Data Cleaning Pipeline
• ML Optimization/Testing

Feb 2025 - Current

This work is still ongoing.

The thesis aims to explore some different ways to build data streaming pipelines in the cloud. I develop an end-to-end streaming pipeline that uses real-time stock market data and derives analytics from it. The data is ingested with Kafka, processed with Spark/Databricks, and visualized with PowerBI. All services are built on scalable cloud compute that is able to handle massive volumes of data.

I have learned so far:

• AWS
• Databricks
• Spark
• ETL/ELT Pipelines
• Data Cleaning/Validation
• Medallion Architecture

Jan 2024 - May 2024

In this project our team developed a building façade recognition algorithm based on semantic segmentation. Multiple different methodologies needed to be implemented and combined into one fast algorithm to make it robust against visual variations caused by weather, time of day, seasons, and foliage. Additionally, the algorithm had to be as efficient as possible to be able to run it seamlessly on a mobile device.

The structure of the algorithm consists of the following filtering methodologies: number of stories, feature counting, wall/door color, text matching, roof shape, and wall texture. I specialized to the texture analysis and developed a Local Binary Pattern algorithm that filters images based on texture similarity. The results were deemed satisfactory by the customer. More detais can be seen in the project report.

I learned:

• Strong Problem Solving
• Understanding Customer Needs
• OpenCV Computer Vision
• Local Binary Pattern (LBP) Algorithm
• Scrum

Jan 2023 - May 2023

The app tracks users' mobile device and prevents them from getting lost in Finnish forests while picking berries. Users can see their own and friends' routes. If internet connection is lost, you and your friends can see each others' latest logged location on the map.

Our team used scrum and agile software development methodologies to facilitate the development of the product. It was essential to understand the customer's needs well enough to be able to adapt and find the best possible solutions for each problem without the need to continuously bother the customer with overly technical details and questions.

Our team resumed the work of a previous development team. The full stack development consisted of Python for backend, and TypeScript for frontend. The app is built using Reach Native and Expo, and we were able set it up for every popular mobile operating system. I'm proud of our solutions for many problems regarding the compatibility of dependencies and other technologies used, such as React Native, Expo, map tiles, and caching.

I learned:

• Deploying the app to the university server
• A working test version on a real phone (in addition to emulators)
• Code refactoring for quality
• A script that enables a fast setup for a new developer to begin the development in a new environment
• Resolving Dependency Conflicts
• All user data is encrypted and safe in the database
• In-app functionality to add, track, and discard multiple friends at the same time (+ settings and UI for it)
• The map source can be changed
• Ability to switch language
• UI overhaul
• Various quality of life features for users and developers

Mar 2022 - Apr 2022

A bare-bones forum website similar to Reddit, where users can discuss topics. Users can create accounts and post, like, and comment threads and other comments. Users can also send private messages to each other and search other users.

I learned:

• Frontend & Backend software architectures and development
• Python
• SQL
• Web Development
• Testing
• Databases
• Software Architecture

Oct 2023 - Dec 2023

In this project, our task was to predict the saturation vapor pressure of different molecules based on their chemical and physical features. To achieve this, we experimented with multiple different machine learning models and different techniques to find and train the best possible model for the task. The project report contains a demonstration of exploratory data analysis, principal component analysis, different ML models, hyperparameter optimization, feature selection, and our conclusions to all of the above.

The course staff organized a Kaggle competition regarding the project (link below), where each group posted their results. Our group performed averagely. Our group's R-squared score was 0.6729 while the winning group had a score of 0.7275.

I learned:

• Exploratory Data Analysis (EDA)
• Feature Selection
• Hyperparameter Optimization
• Data Cleaning Pipeline
• Linear Regression
• Decision Trees
• Random Forests
• Gradient Boosting

May 2023 - Jun 2023

This app clusters similar texts into groups. It is self-made from the ground up, using no premade algorithms from external libraries. The app uses a dataset of old BBC news articles and takes additional texts as input from the user. It sorts those articles into groups based on text similarity using TF-IDF matrix and K-means clustering algorithm.

I learned:

• K-Means Unsupervised Machine Learning Algorithm (With TF-IDF Matrices)
• Data Cleaning Pipeline
• Testing
• Software Architecture