2023-2024 Capstone Projects

As the necessary tests for the preliminary assessment of drug carcinogenicity are time-consuming and expensive, the methods to predict their results using machine learning models become more and more popular. However, it is often the case that the dataset with labeled molecules is small. We try to address the problem by training the model on several tasks simultaneously using the one-primary multiple-auxiliaries approach which enables us to leverage other datasets besides the one for the primary task thus solving the problem of small datasets. We show that the results for in-vivo rat carcinogenicity improve when using carefully chosen auxiliary tasks.

The main goal of this project is to help the deaf community or people with hearing impairment to live a more productive and safe life. For solving that important problem this project focuses on two specific types of environmental sounds: road noises and emergency sirens. These are the main types of environmental sounds which are needed to warn people of potential dangers. This project uses unsupervised machine learning models to analyze and classify those sounds. The selection of the model was made based on the experiments on available data. Since the data doesn’t distinguish different types of road noise or emergency sirens, we used a clustering technique to group similar sounds without prior knowledge of their specific categories. The final model can distinguish between different types of road noise and emergency vehicle sirens by detecting patterns in the data. The practical part of this research is embodied in a prototype glove. This glove converts recognized sounds into vibrations. The intensity and pattern of these vibrations vary by decibel. By translating auditory signals into tactile ones, the glove helps deaf people better understand important sounds in their environment. There is also a screen on the glove that shows the type and group of the given sound.

The rapid proliferation of online news sources has led to a wealth of information that can be analyzed to gain insights into public sentiment and opinions about various entities and topics. This paper aims to explore and harness the capabilities of  OpenAI’s GPT-3.5, a large language model (LLM), as an easily accessible text processing component, specifically for sentiment analysis. We show the advantages and disadvantages of integrating an LLM as a component, we look for relationships among entities and build a methodologically simple sentiment estimator to infer the sentiment of all the entities within the scraped dataset.

Keywords: News.am, Tert.am, Armenpress, news, sentiment analysis, LLM

In the rapidly developing sphere of iGaming, sustaining customer loyalty is crucial for the success of companies in this dynamic industry. The purpose of this Capstone project is to enhance the Customer Relationship Management (CRM) of one of the Armenian gaming companies with an innovative approach by automating the Customer Retention team’s workload processes. The current manual CRM practices often result in inefficiencies, errors and delays in addressing customer needs, leading to potential customer dissatisfaction and churn. The aim of this project is to create an automated system that will optimize CRM retention workload and make it more efficient and effective. The system will help the company to increase customer satisfaction as it will reduce manual intervention and enable employees to provide timely, personalized interactions to customer preferences, behaviors and engagement.

This paper suggests a customized approach for bookstore recommendation algorithms using Zangak Bookstore as a case study. With the help of content-based filtering and modern natural language processing algorithms, our models generate personalized book recommendations based on customer interests, past experiences, and book characteristics. The process uses content-based filtering to create item representations based on attributes, including titles, authors, languages, and descriptions. We use the KeyBert model to extract essential keywords from translated descriptions to improve the quality of recommendations. Based on the book’s features, the paper presents two primary models mentioned in Chapter 4. Subsequently, the models undergo expansion to consider customers’ preferences based on their past purchases. Eventually, the resulting extended model integrates the two previous methodologies. Our methodology presents a fresh approach to tailored book recommendations for local bookstores in Armenia, improving customers’ browsing and buying experiences, specifically at Zangak Bookstore, and possibly contributing to the broader e-commerce scene.

Keywords: Recommendation systems, Books, Zangak Bookstore, Content-based filtering, KeyBert, BERT, Translation, Keyword Extraction

This paper details a capstone project that leverages OpenAI APIs and text processing tools to automate timeline generation in investment arbitration using data from the FDIMOOT competition. The project integrates multiple OpenAI APIs, such as GPT-3.5 Turbo and GPT-4 Turbo with vision capabilities, to extract, clean, and analyze textual data from documents, converting them into a structured timeline of events. This methodology addresses the complexities and inefficiencies associated with manual timeline construction in legal documents. Key outcomes demonstrate the potential of OpenAI-driven processes to enhance accuracy and efficiency in legal data analysis. This paper explores the methods used, discusses the results, and acknowledges limitations while proposing future research directions.

Index Terms— Timeline Generation, OpenAI APIs, Document Analysis

Utilizing machine learning algorithms has proven to be a potent method for managing risk and preventing financial fraud. However, the requirement for coding expertise and deep knowledge of machine learning stands as a barrier for users who need such skills. This study offers an innovative, user-friendly framework that uses machine learning to address this difficulty. The paper studies the performance of three machine learning models, Random Forest, SVM, and XGboost, to classify, detect, and predict fraudulent transactions.

Keywords: Fraud detection, machine learning, risk management tool, data cleaning, data visualization, classification, machine learning algorithms, streamlit, random forest, XGboost, SVM, finance, detection models, dynamic
environments.

Composing music is traditionally a realm of human creativity, intuition, and emotion. However, recent advancements in deep learning algorithms have opened directions for the exploration of automated music generation. The primary objective of this research is to create a generative AI model capable of producing moderate-sounding compositions that resonate with the cultural nuances of Armenia. We seek to expand the realm of creative possibilities in Armenian music composition and to contribute to the preservation and promotion of Armenian cultural heritage. We employ the MusicGen model for text-to-music generation with efficient implementation of the training pipeline, significantly enhancing both speed and memory utilization. The MusicGen small checkpoint is fine-tuned on a musical corpus of 62.8 hours of Armenian music extracted from the internet. The results were satisfactory where the human evaluators provided on average a rating of 3.84 points from 5 regarding the generated music quality and 3.96 points from 5 concerning the relevance of the given text condition to the generation, indicating the model’s ability to generate adequate sounding Armenian compositions aligning with the textual descriptions provided to the model. By pioneering the development of an advanced music generation model for Armenian music, we not only showcase the potential of artificial intelligence in creative domains but also foster Armenian cultural appreciation and innovation.

Keywords: MusicGen, Generative AI, Armenian Music Generation, Transformers, Music.

The main objective of this capstone work is to create and implement a car price prediction system based on the analysis of information from the leading site for selling cars in Armenia. The project’s problem lies in the high volatility of prices, and data is non-standardized, making it hard to buy and sell cars in Armenia. We propose an approach based on machine learning techniques for analyzing and processing collected data. To achieve this, two pipelines were developed and tested for the study: one without car images and another using the Resnet18 model, deep learning technology for photo analysis. Interestingly, models trained with image features yielded similar accuracy levels on predictions. Henceforth, an online platform was built on Flask with features enabling users to evaluate their vehicle worth, such as an interactive user interface for entering car attributes and getting price estimates and a statistical graphics page including market analytics. This project contributes to improving the transparency and efficiency of the Armenian car market and opens up new directions for further research in machine learning in the automotive industry.

Keywords: Car Price Forecasting System, Armenian automobiles, Deep learning technologies, Photo Analysis

In recent years, automatic speech recognition (ASR) systems have significantly improved, especially in languages with a vast amount of transcribed speech data. However, ASR systems tend to perform poorly for low-resource languages with fewer resources, such as minority and regional languages. This study introduces a novel pipeline designed to generate ASR training datasets from audiobooks, which typically feature a single transcript associated with hours-long audios. The common structure of these audiobooks poses a unique challenge due to the extensive length of audio segments, whereas optimal ASR training requires segments ranging from 4 to 15 seconds. To address this, we propose a method for effectively aligning audio with its corresponding text and segmenting it into lengths suitable for ASR training. Our approach simplifies data preparation for ASR systems in low-resource languages and demonstrates its application through a case study involving the Armenian language. Our method, which is "portable" to many low-resource languages, not only mitigates the issue of data scarcity but also enhances the performance of ASR models for underrepresented languages.

Keywords: ASR, low-resource languages, dataset creation, alignment.

My project is aimed to create a recommendation system for hotels in Rome. This will create an opportunity to just chat with the AI assistant and give specific details about what kind of hotel one needs to attend mentioning features like location, street, accommodations, rooms, number of people, purpose, feedback and so on and get the best match for the mentioned needs. The data I’m using is an officially licensed dataset from Airbnb which is appropriate to use for the recommendation system. I used embeddings for each row in the dataset, which gives the opportunity to get the answers for the specific question regarding all the information about the hotels in Rome in a few seconds. Using various techniques that we’ll discuss later I have created an interactive chatbot specialized in hotel recommendations which is easy to use just by transferring your thoughts as an input.

The appearance of cryptocurrencies in financial markets has changed how people think about digital assets and potential variants for investment. This paper aims to explore the implementation of machine learning models to predict cryptocurrency price movements. The BTCUSDT pair has been chosen for the case study. The research methodology includes a review of several papers about machine learning models being used for recognizing patterns in financial markets, the collection of trading data, the preparation of the data for analysis, and the application of ML models that capture dependencies and patterns in price movements. This work aims to contribute to the growing field of financial technology by offering a comprehensive analysis of ML models in cryptocurrency prediction and reliable forecasting tools in the digital finance landscape.

Keywords: cryptocurrency - digital currency, BTCUSDT - the pair of bitcoin and USD tokens, ML - machine learning

This capstone project looks at incorporating artificial intelligence in the writing training, emphasizing the execution and examination of the provision of advanced language models, particularly GPT-2 and GPT-J-6B. The project’s primary purpose was to refine the AI writing apparatus to get the most out of them. The project would be achieved by fine-tuning the models using Project Gutenberg, which has a lot of literary data. The study proposed to uphold the assertion that fine-tuned models will produce a better text in both coherent and contextual appropriateness and when one is required to write in modes such as creative, factual, and technical writing. Contrary to the predictions, the data analysis was done, which showed that the tuned models frequently failed to work better than the default configurations. This suboptimal prediction was caused by problems such as overfitting to the literary styles of the training data, which led to the loss of generalization property and, thus, to poor performance of the models on tasks involving a wide range of tasks and different data. The project also tackled the technical difficulties connected with optimizing the large models of the language and many other areas, such as computational resource management and ethical issues in deployment. The result underlines the complicatedness of establishing AI in writing assistance, raising the necessity of maintaining a balance between the profiling of specialists for individual cases and the covering of a large territory of specialization. Our study is, therefore, another contribution whose purpose is to articulate how AI should contribute to keeping the authenticity of writing with human creativity. At this moment, the research’s primary results will force the interaction process of AI-funded writing in novel directions; instead it will direct future planning for new AI modifications, which will be aimed at practical, ethic and inclusive implementations.