About

Dr. Maryan Rizinski joined Boston University’s Metropolitan College as a full-time associate professor of the practice in computer science in 2025. He holds a PhD in computer science, as well as MS and BS degrees in electrical engineering and information technologies from the University Ss. Cyril and Methodius in Skopje. His interdisciplinary research bridges real-world applications with novel approaches in machine learning, natural language processing, and data analytics, with a focus on ethical, transparent, and cost-effective AI solutions in the financial industry and related domains. His work advances decision-making, risk mitigation, and the interpretability of deep-learning models. With over fifteen years of industry experience, Dr. Rizinski has led software engineering teams and managed complex projects across various domains including automotive, smart buildings, and consumer goods. He has held leadership roles at Bosch Digital, where he delivered industry-focused solutions for multinational corporations, and has advised venture capital firms and mentored startups on strategy, IT innovation, and risk mitigation. Since 2013, he has also been teaching computer science courses at BU MET, demonstrating a strong record in online teaching, course development, and facilitation. He was awarded the BU MET 2023 Roger Deveau Memorial Part-Time Faculty Award for Excellence in Teaching.

Research topics

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Selected publications

• Forensic Output Scale (FOS): Dataset of 79 Deepfake Detection Methods Classified by Forensic Output Type

  Zenodo (CERN European Organization for Nuclear Research) · 2026-05-14

  datasetOpen access

  This dataset accompanies the paper "Detecting Deepfakes for the Courtroom: A Multi-Modal Forensic Output Framework" (EAI CSECS 2026). It catalogs 79 deepfake detection methods across image, video, audio-visual, and audio modalities, classified by the Forensic Output Scale (FOS), a five-level taxonomy of detector output types: L0 Detection (score/probability), L1 Localization (where), L2 Attribution (what generator/method), L3 Narration (natural-language explanation), and L4 Verification (a specific, juror-checkable factual claim). The dataset reveals that 57% of surveyed methods (45/79) ship only L0 outputs, while no method currently produces L4 verification output. Seven methods compute L4-capable internal signals (mouth temporal anomalies, action unit relationships, blood-flow rPPG, lip-audio alignment, pupil geometry, etc.) but collapse them to scalar scores at inference — a gap we term the "L4 readiness gap." Contents (10 sheets): README — taxonomy definitions and dataset documentation Methods — master catalog of all 79 methods with year, venue, modality, FOS level, output type, internal signal (if L4-capable), notes, and Chicago author-date reference L0 Detection / L1 Localization / L2 Attribution / L3 Narration / L4 Verification — methods split by FOS level L4-Capable (ships L0) — the seven detectors with internal verifiable signals plus the L4 claim each would produce if surfaced and how a juror could verify it Datasets — 21 evaluation datasets supporting explainable deepfake detection Summary — distribution statistics across FOS levels and modalities Intended use: Forensic practitioners selecting court-admissible detectors, researchers identifying gaps in explainability, and policy/legal scholars assessing detector suitability under Daubert and analogous evidence standards.

  PublisherDOI

• When AI Fails, What Works? A Data-Driven Taxonomy of Real-World AI Risk Mitigation Strategies

  Open MIND · 2026-03-04

  preprint

  Large language models (LLMs) are increasingly embedded in high-stakes workflows, where failures propagate beyond isolated model errors into systemic breakdowns that can lead to legal exposure, reputational damage, and material financial losses. Building on this shift from model-centric risks to end-to-end system vulnerabilities, we analyze real-world AI incident reporting and mitigation actions to derive an empirically grounded taxonomy that links failure dynamics to actionable interventions. Using a unified corpus of 9,705 media-reported AI incident articles, we extract explicit mitigation actions from 6,893 texts via structured prompting and then systematically classify responses to extend MIT's AI Risk Mitigation Taxonomy. Our taxonomy introduces four new mitigation categories, including 1) Corrective and Restrictive Actions, 2) Legal/Regulatory and Enforcement Actions, 3) Financial, Economic, and Market Controls, and 4) Avoidance and Denial, capturing response patterns that are becoming increasingly prevalent as AI deployment and regulation evolve. Quantitatively, we label the mitigation dataset with 32 distinct labels, producing 23,994 label assignments; 9,629 of these reflect previously unseen mitigation patterns, yielding a 67% increase of the original subcategory coverage and substantially enhancing the taxonomy's applicability to emerging systemic failure modes. By structuring incident responses, the paper strengthens "diagnosis-to-prescription" guidance and advances continuous, taxonomy-aligned post-deployment monitoring to prevent cascading incidents and downstream impact.

  DOI

• Forensic Output Scale (FOS): Dataset of 79 Deepfake Detection Methods Classified by Forensic Output Type

  Zenodo (CERN European Organization for Nuclear Research) · 2026-05-15

  datasetOpen access

  This dataset accompanies the paper "Detecting Deepfakes for the Courtroom: A Multi-Modal Forensic Output Framework" (EAI CSECS 2026). It catalogs 79 deepfake detection methods across image, video, audio-visual, and audio modalities, classified by the Forensic Output Scale (FOS), a five-level taxonomy of detector output types: L0 Detection (score/probability), L1 Localization (where), L2 Attribution (what generator/method), L3 Narration (natural-language explanation), and L4 Verification (a specific, juror-checkable factual claim). The dataset reveals that 57% of surveyed methods (45/79) ship only L0 outputs, while no method currently produces L4 verification output. Seven methods compute L4-capable internal signals (mouth temporal anomalies, action unit relationships, blood-flow rPPG, lip-audio alignment, pupil geometry, etc.) but collapse them to scalar scores at inference — a gap we term the "L4 readiness gap." Contents (10 sheets): README — taxonomy definitions and dataset documentation Methods — master catalog of all 79 methods with year, venue, modality, FOS level, output type, internal signal (if L4-capable), notes, and Chicago author-date reference L0 Detection / L1 Localization / L2 Attribution / L3 Narration / L4 Verification — methods split by FOS level L4-Capable (ships L0) — the seven detectors with internal verifiable signals plus the L4 claim each would produce if surfaced and how a juror could verify it Datasets — 21 evaluation datasets supporting explainable deepfake detection Summary — distribution statistics across FOS levels and modalities Intended use: Forensic practitioners selecting court-admissible detectors, researchers identifying gaps in explainability, and policy/legal scholars assessing detector suitability under Daubert and analogous evidence standards.

  PublisherDOI

• When AI Fails, What Works? A Data-Driven Taxonomy of Real-World AI Risk Mitigation Strategies

  ArXiv.org · 2026-03-04

  articleOpen access

  Large language models (LLMs) are increasingly embedded in high-stakes workflows, where failures propagate beyond isolated model errors into systemic breakdowns that can lead to legal exposure, reputational damage, and material financial losses. Building on this shift from model-centric risks to end-to-end system vulnerabilities, we analyze real-world AI incident reporting and mitigation actions to derive an empirically grounded taxonomy that links failure dynamics to actionable interventions. Using a unified corpus of 9,705 media-reported AI incident articles, we extract explicit mitigation actions from 6,893 texts via structured prompting and then systematically classify responses to extend MIT's AI Risk Mitigation Taxonomy. Our taxonomy introduces four new mitigation categories, including 1) Corrective and Restrictive Actions, 2) Legal/Regulatory and Enforcement Actions, 3) Financial, Economic, and Market Controls, and 4) Avoidance and Denial, capturing response patterns that are becoming increasingly prevalent as AI deployment and regulation evolve. Quantitatively, we label the mitigation dataset with 32 distinct labels, producing 23,994 label assignments; 9,629 of these reflect previously unseen mitigation patterns, yielding a 67% increase of the original subcategory coverage and substantially enhancing the taxonomy's applicability to emerging systemic failure modes. By structuring incident responses, the paper strengthens "diagnosis-to-prescription" guidance and advances continuous, taxonomy-aligned post-deployment monitoring to prevent cascading incidents and downstream impact.

  PublisherOA PDF

• Forensic Output Scale (FOS): Dataset of 79 Deepfake Detection Methods Classified by Forensic Output Type

  Zenodo (CERN European Organization for Nuclear Research) · 2026-05-14

  datasetOpen access

  This dataset accompanies the paper "Detecting Deepfakes for the Courtroom: A Multi-Modal Forensic Output Framework" (EAI CSECS 2026). It catalogs 79 deepfake detection methods across image, video, audio-visual, and audio modalities, classified by the Forensic Output Scale (FOS), a five-level taxonomy of detector output types: L0 Detection (score/probability), L1 Localization (where), L2 Attribution (what generator/method), L3 Narration (natural-language explanation), and L4 Verification (a specific, juror-checkable factual claim). The dataset reveals that 57% of surveyed methods (45/79) ship only L0 outputs, while no method currently produces L4 verification output. Seven methods compute L4-capable internal signals (mouth temporal anomalies, action unit relationships, blood-flow rPPG, lip-audio alignment, pupil geometry, etc.) but collapse them to scalar scores at inference — a gap we term the "L4 readiness gap." Contents (10 sheets): README — taxonomy definitions and dataset documentation Methods — master catalog of all 79 methods with year, venue, modality, FOS level, output type, internal signal (if L4-capable), notes, and Chicago author-date reference L0 Detection / L1 Localization / L2 Attribution / L3 Narration / L4 Verification — methods split by FOS level L4-Capable (ships L0) — the seven detectors with internal verifiable signals plus the L4 claim each would produce if surfaced and how a juror could verify it Datasets — 21 evaluation datasets supporting explainable deepfake detection Summary — distribution statistics across FOS levels and modalities Intended use: Forensic practitioners selecting court-admissible detectors, researchers identifying gaps in explainability, and policy/legal scholars assessing detector suitability under Daubert and analogous evidence standards.

  PublisherDOI

• PLMN-GraphSim dataset for "Beyond 5G Architectural Constraints: Designing Scalable-by-Design User Planes for Mobile Networks via 6G-RUPA"

  CORA.Repositori de Dades de Recerca · 2026-04-17

  datasetOpen access

  <p>This dataset contains the processed geospatial and network topology inputs required to reproduce the experiments presented in the paper <em>"Beyond 5G Architectural Constraints: Designing Scalable-by-Design User Planes for Mobile Networks via 6G-RUPA"</em>.</p> <ol> <li><strong>Spain Scenario Data</strong> <ul> <li><strong>Demographics:</strong> Standardized municipality data (boundaries, population) derived from the National Statistics Institute (INE).</li> <li><strong>Network Topology:</strong> Real-world cellular tower (gNB) locations for major operators (MCC 214), filtered and processed from OpenCellID.</li> </ul> </li> <li><strong>USA Scenario Data</strong> <ul> <li><strong>Demographics:</strong> High-resolution census tract and population data derived from the US Census Bureau databases.</li> <li><strong>Network Topology:</strong> Cellular infrastructure density and locations derived from OpenCellID.</li> </ul> </li> </ol> <h3>Structure</h3> <p>The data is organized by country directory (<code>/spain</code>, <code>/usa</code>) and formatted for the simulation environment:</p> <ul> <li><code>municipalities.csv</code>: Contains population counts and centroid coordinates.</li> <li><code>regions.geojson</code>: Contains geometric boundaries for spatial indexing and user distribution.</li> <li><code>opencellid/*.csv</code>: Contains processed base station coordinates.</li> </ul> <h3>Usage</h3> <p>This dataset is intended to be extracted into the <code>data</code> directory of the simulation framework. Full processing scripts used to generate these files from the raw sources are available in the associated code repository.</p>

  PublisherDOI

• Rethinking Inequality: The Complex Dynamics Beyond the Kuznets Curve

  Data · 2025-06-14 · 3 citations

  articleOpen access

  Income inequality has emerged as a defining challenge of our time, particularly in advanced economies, where the gap between rich and poor has reached unprecedented levels. This study analyzes income inequality trends from 2000 to 2023 across developed countries (the United States, the United Kingdom, Germany, and France) and developing nations using World Bank Gini coefficient data. We employ comprehensive visualization techniques, Pareto distribution analysis, and ARIMA time-series forecasting models to evaluate the effectiveness of the Kuznets curve as a predictor of income inequality. Our analysis reveals significant deviations from the traditional inverse U-shaped Kuznets curve across all examined countries, with persistent volatility rather than the predicted decline in inequality. Forecasts using ARIMA and neural networks indicate continued fluctuations in inequality through 2030, with the U.S. and Germany showing upward trends while France and the UK demonstrate relative stability. These findings challenge the conventional Kuznets hypothesis and demonstrate that contemporary inequality patterns are influenced by factors beyond economic development, including technological change, globalization, and policy choices. This research contributes to the literature by providing empirical evidence that the Kuznets curve has limited predictive power in modern economies, informing policymakers about the need for targeted interventions to address persistent inequality rather than relying on economic growth alone.

  PublisherOA PDFDOI

• AI Agents in Finance and Fintech: A Scientific Review of Agent-Based Systems, Applications, and Future Horizons

  Computers, materials & continua/Computers, materials & continua (Print) · 2025-10-30 · 2 citations

  articleOpen access1st authorCorresponding

  Artificial intelligence (AI) is reshaping financial systems and services, as intelligent AI agents increasingly form the foundation of autonomous, goal-driven systems capable of reasoning, learning, and action. This review synthesizes recent research and developments in the application of AI agents across core financial domains. Specifically, it covers the deployment of agent-based AI in algorithmic trading, fraud detection, credit risk assessment, robo-advisory, and regulatory compliance (RegTech). The review focuses on advanced agent-based methodologies, including reinforcement learning, multi-agent systems, and autonomous decision-making frameworks, particularly those leveraging large language models (LLMs), contrasting these with traditional AI or purely statistical models. Our primary goals are to consolidate current knowledge, identify significant trends and architectural approaches, review the practical efficiency and impact of current applications, and delineate key challenges and promising future research directions. The increasing sophistication of AI agents offers unprecedented opportunities for innovation in finance, yet presents complex technical, ethical, and regulatory challenges that demand careful consideration and proactive strategies. This review aims to provide a comprehensive understanding of this rapidly evolving landscape, highlighting the role of agent-based AI in the ongoing transformation of the financial industry, and is intended to serve financial institutions, regulators, investors, analysts, researchers, and other key stakeholders in the financial ecosystem.

  PublisherDOI

• Evaluating Trustworthiness in AI: Risks, Metrics, and Applications Across Industries

  Electronics · 2025-07-04 · 39 citations

  articleOpen accessCorresponding

  Ensuring the trustworthiness of artificial intelligence (AI) systems is critical as they become increasingly integrated into domains like healthcare, finance, and public administration. This paper explores frameworks and metrics for evaluating AI trustworthiness, focusing on key principles such as fairness, transparency, privacy, and security. This study is guided by two central questions: how can trust in AI systems be systematically measured across the AI lifecycle, and what are the trade-offs involved when optimizing for different trustworthiness dimensions? By examining frameworks such as the NIST AI Risk Management Framework (AI RMF), the AI Trust Framework and Maturity Model (AI-TMM), and ISO/IEC standards, this study bridges theoretical insights with practical applications. We identify major risks across the AI lifecycle stages and outline various metrics to address challenges in system reliability, bias mitigation, and model explainability. This study includes a comparative analysis of existing standards and their application across industries to illustrate their effectiveness. Real-world case studies, including applications in healthcare, financial services, and autonomous systems, demonstrate approaches to applying trust metrics. The findings reveal that achieving trustworthiness involves navigating trade-offs between competing metrics, such as fairness versus efficiency or privacy versus transparency, and emphasizes the importance of interdisciplinary collaboration for robust AI governance. Emerging trends suggest the need for adaptive frameworks for AI trustworthiness that evolve alongside advancements in AI technologies. This paper contributes to the field by proposing a comprehensive review of existing frameworks with guidelines for building resilient, ethical, and transparent AI systems, ensuring their alignment with regulatory requirements and societal expectations.

  PublisherOA PDFDOI

• Sentiment Analysis in Finance: From Transformers Back to eXplainable Lexicons (XLex)

  IEEE Access · 2024-01-01 · 33 citations

  articleOpen access1st authorCorresponding

  Lexicon-based sentiment analysis in finance leverages specialized, manually annotated lexicons created by human experts to effectively extract sentiment from financial texts. Although lexicon-based methods are simple to implement and fast to operate on textual data, they require considerable manual annotation efforts to create, maintain, and update the lexicons. These methods are also considered inferior to the deep learning-based approaches, such as transformer models, which have become dominant in various natural language processing (NLP) tasks due to their remarkable performance. However, their efficacy comes at a cost: these models require extensive data and computational resources for both training and testing. Additionally, they involve significant prediction times, making them unsuitable for real-time production environments or systems with limited processing capabilities. In this paper, we introduce a novel methodology named eXplainable Lexicons (XLex) that combines the advantages of both lexicon-based methods and transformer models. We propose an approach that utilizes transformers and SHapley Additive exPlanations (SHAP) for explainability to automatically learn financial lexicons. Our study presents four main contributions. Firstly, we demonstrate that transformer-aided explainable lexicons can enhance the vocabulary coverage of the benchmark Loughran-McDonald (LM) lexicon. This enhancement leads to a significant reduction in the need for human involvement in the process of annotating, maintaining, and updating the lexicons. Secondly, we show that the resulting lexicon outperforms the standard LM lexicon in sentiment analysis of financial datasets. Our experiments show that XLex outperforms LM when applied to general financial texts, resulting in enhanced word coverage and an overall increase in classification accuracy by 0.431. Furthermore, by employing XLex to extend LM, we create a combined dictionary, XLex+LM, which achieves an even higher accuracy improvement of 0.450. Thirdly, we illustrate that the lexicon-based approach is significantly more efficient in terms of model speed and size compared to transformers. Lastly, the proposed XLex approach is inherently more interpretable than transformer models. This interpretability is advantageous as lexicon models rely on predefined rules, unlike transformers, which have complex inner workings. The interpretability of the models allows for better understanding and insights into the results of sentiment analysis, making the XLex approach a valuable tool for financial decision-making.

  PublisherOA PDFDOI

Frequent coauthors

• Dimitar Trajanov

  Boston University

  15 shared

• Hristijan Peshov

  Saints Cyril and Methodius University of Skopje

  4 shared

• Kostadin Mishev

  Saints Cyril and Methodius University of Skopje

  4 shared

• Ljupčo Kocarev

  Saints Cyril and Methodius University of Skopje

  4 shared

• Eugene Pinsky

  Boston University

  3 shared

• Vignesh Sankaradas

  Boston University

  3 shared

• Andrej Jankov

  Saints Cyril and Methodius University of Skopje

  3 shared

• Milos Jovanovik

  Saints Cyril and Methodius University of Skopje

  3 shared

Awards & honors

• 2023 Roger Deveau Memorial Part-Time Faculty Award for Excel…