Artificial intelligence-based system for real-time credit rating of small and medium-sized enterprises

Abstract

A system for real-time creditworthiness assessment of small and medium-sized enterprises (SMEs) based on agent-based artificial intelligence (AI), the system includes: a large number of autonomous agents connected to each other via secure communication protocols; a data ingestion processing unit configured to capture and preprocess heterogeneous data streams from structured and unstructured sources, including APIs for financial transactions, tax databases, e-invoicing registers, inventory management systems, social sentiment feeds, and IoT-enabled devices, with preprocessing including schema harmonization, anomaly detection, and encryption-based integrity preservation; an ensemble of scoring control units configured to generate dynamic credit scores, wherein the ensemble includes machine learning models, including gradient-enhanced decision trees, recurrent neural networks, and graph neural networks, the graph neural networks being configured to assess relational dependencies between SMEs, suppliers, and customer nodes in a business ecosystem; an audit controller configured to generate regulatory-compliant justification paths for each credit score; a feedback processing unit configured to update the evaluation control unit ensemble using reinforcement learning signals derived from observed repayment behavior, payment default trends, and fraud detection markers; and a user interface configured to display timestamped credit histories, confidence intervals and explanatory justifications, with the entire system running on a containerized cloud-native infrastructure with blockchain-based ledgers for version control of each point update.

Description

AREA OF INVENTION

[0001] The present invention relates to technologies for artificial intelligence, financial analysis, and risk assessment. In particular, the invention relates to an agent-based AI framework that autonomously assesses, updates, and explains the creditworthiness of small and medium-sized enterprises (SMEs) using distributed multi-agent architectures, explainable AI methods, and dynamic, real-time streaming data integration. The invention also includes a physical machine and device structure configured for use in financial institutions, credit bureaus, and business ecosystems. BACKGROUND OF THE INVENTION

[0002] Small and medium-sized enterprises (SMEs) form the backbone of global economic growth and job creation. However, their access to credit is significantly hampered by the inadequacy of traditional valuation systems, which rely primarily on static financial statements, outdated historical records, and rigid rule-based models. Such conventional systems fail to capture the dynamic operational dynamics of SMEs, where sudden changes in the supply chain, customer base, or market sentiment can drastically alter risk.

[0003] Existing machine learning-based credit scoring platforms, while more sophisticated than traditional models, still have limitations. For example, earlier systems apply machine learning to transaction data but lack the necessary agent-based orchestration and contextual inference. Therefore, they are unsuitable for SMEs with limited credit histories. Similarly, US20200236238A1 assesses creditworthiness using natural language processing and behavioral data but is geared more towards individuals than SMEs. Other developments, such as US20210143599A1 and US20200341765A1, focus either on consumer loans or portfolio risks and do not offer decentralized, agent-based, real-time updates.

[0004] Therefore, there is an urgent need for a system that utilizes agent-based AI, distributed coordination and explainable results, and is capable of adaptively monitoring the creditworthiness of SMEs in real time and providing transparent, regulatory-compliant justifications.

[0005] Small and medium-sized enterprises (SMEs) form the backbone of the global economy. They provide the majority of jobs and contribute significantly to gross domestic product in both industrialized and developing countries. Despite their undeniable importance, SMEs repeatedly struggle to obtain formal loans, primarily due to shortcomings in conventional credit assessment methods. The technological foundation of this invention therefore lies at the intersection of artificial intelligence, financial risk assessment, and adaptive, data-driven decision-making. The development of agile, real-time, and transparent credit assessment systems is urgently needed. To understand this need, it is crucial to examine existing credit assessment solutions, their methodologies, and the associated drawbacks that render them unsuitable for dynamically evolving SME ecosystems.

[0006] Traditional credit scoring systems were originally developed for large corporations and later extended to consumer loans. These systems rely heavily on static financial statements such as balance sheets, profit and loss statements, collateral declarations, and—in the case of individuals—credit bureaus and repayment histories. However, for SMEs, the availability of structured and consistent historical financial data is often limited. Many SMEs are family businesses or operate with informal accounting structures, meaning their records are sparse or incomplete. Consequently, traditional credit scoring tends to underestimate the creditworthiness of SMEs, treating them as high-risk by default, regardless of their actual operational health.This rigidity systematically excludes potentially viable businesses from formal credit channels, stifling growth and innovation. Furthermore, these models are typically updated quarterly or annually and fail to capture sudden but critical changes in an SME's operational dynamics, such as a surge in demand, the acquisition of a new supplier, or an unexpected disruption in cash flow.

[0007] In response to the limitations of purely rule-based systems, machine learning and predictive analytics have been integrated into credit risk assessment over the past decade. An example of this is the state of the art, such as US20190082316A1, which describes an automated credit risk assessment system that applies machine learning to financial transaction data. While this approach represents an improvement in terms of adaptability and predictive power, it still has significant limitations. The models described in such solutions rely on large volumes of structured data and long histories of transaction information, which are often unavailable to SMEs. In scenarios with inconsistent transaction records or those limited to a few years, the system tends to generate distorted or unstable credit forecasts.Furthermore, the lack of multi-agent orchestration and contextual thinking prevents these systems from understanding the differentiated realities of SMEs operating in a volatile market environment.

[0008] Another strand of existing solutions focuses on behavioral analytics and natural language processing. For example, US20200236238A1 emphasizes the use of non-traditional data sources, including text-based sentiment and behavioral patterns, to assess creditworthiness. While such techniques extend the scope beyond financial statements, they are geared more toward individuals and consumer loans than SMEs. SME operations involve multifaceted interactions—supply chains, supplier reliability, employee performance, customer acquisition, and tax compliance—that cannot be adequately captured by behavioral or language data alone. Moreover, models based on such inputs often lack transparency. They generate statistically valid but difficult-to-interpret credit scores.This makes them less suitable for financial institutions that must meet regulatory requirements that demand transparent decision-making processes.

[0009] Further developments in AI-based creditworthiness assessment, as described in US20210143599A1, are based on engines specifically designed for consumer loans. These systems utilize sophisticated forecasting techniques but fail to consider the decentralized, collaborative structures in which SMEs operate. An SME's creditworthiness reflects not only its internal balance sheet but is also closely linked to its suppliers, customers, and regional economic conditions. Without mechanisms to account for relational dependencies between SME networks, these models are inherently narrow. They do not consider credit contagion risks or the opportunities arising from collaboration between SMEs. For example, an SME with strong and reliable suppliers may have a lower credit risk than one dependent on unstable or inconsistent partners.However, traditional AI-based engines do not capture this network-based context.

[0010] Some existing risk management systems, such as those described in US20200341765A1, attempt to extend the analysis by focusing on portfolio-level risk. These systems are valuable for banks and large institutions managing diversified investments because they provide insights into risk and volatility at the macro level. However, their usefulness at the individual SME level is limited. They lack the necessary granularity to assess the individual performance of each SME and therefore cannot support adaptive lending decisions on an individual company basis. An SME's operational risk can change dramatically within a few weeks due to new orders, altered input costs, or regulatory changes, but portfolio-level tools obscure these fluctuations in aggregated measurements.

[0011] Machine learning methods in banking, as described in US20210120809A1, attempt to integrate artificial intelligence into credit scoring by training models on large historical datasets. However, these approaches often rely on offline training, which presents significant challenges when applied to SMEs. Offline models are static and cannot adapt in real time to new data inputs such as late payments, sudden regulatory violations, or fraud alerts. SMEs need systems that not only learn continuously but also adapt to evolving environments where new patterns can emerge unpredictably. Without drift detection mechanisms, feedback loops, or reinforcement learning, offline-trained models risk becoming obsolete quickly in volatile SME contexts.

[0012] In addition to these limitations, the lack of explainability is another crucial drawback of existing credit scoring solutions. Financial regulators in all jurisdictions are increasingly demanding that credit decisions, particularly negative ones, be explained to both applicants and regulators. While black-box models are powerful predictors, they lack the transparency necessary for trust and compliance. In the SME sector, where relationships with lenders are often personal and reputation-based, the inability to provide a clear rationale for a score further alienates businesses and fuels distrust of formal financial systems. Explainable AI has been extensively researched in other areas, but its application in credit scoring, especially for SMEs, remains limited in practice.

[0013] Data integration also remains a persistent challenge for existing systems. SMEs generate data from a multitude of sources, including tax systems, invoicing platforms, point-of-sale systems, social media, logistics networks, and even IoT-enabled inventory management systems. Current solutions lack the architectural flexibility to integrate these diverse inputs into coherent credit scoring pipelines. Instead, they typically focus on narrowly defined data types, overlooking valuable signals that could improve the accuracy of credit assessments. Without the aggregation of data from multiple sources, these systems deliver an incomplete and often distorted picture of SME health.

[0014] Furthermore, scalability remains a persistent challenge. Many AI-based systems are resource-intensive, requiring significant computing power and specialized infrastructure to process high-frequency data streams. Smaller financial institutions or microfinance institutions, which play a key role in lending to SMEs, may not be able to deploy such complex systems. Existing models therefore tend to remain the domain of large banks or fintech giants, creating an access gap for smaller lenders who interact directly with SMEs on a large scale.

[0015] In summary, these existing solutions demonstrate that while progress has been made in moving away from rule-based and static credit scoring, none of them adequately address the specific needs of SMEs. They either rely too heavily on historical data, fail to integrate signals from multiple sources in real time, neglect relational and contextual thinking, or lack the explainability and regulatory compliance necessary for sustainable adoption. SMEs inherently operate in volatile, interconnected, and data-poor environments. Therefore, an effective system must be agent-based, adaptive, transparent, and capable of continuous learning while leveraging diverse data ecosystems.The disadvantages of existing solutions thus lead to the technical problem to be solved: enabling real-time, explainable and network-aware credit scoring specifically tailored to SMEs through intelligent agent-based architectures. Objectives of the invention

[0016] The main objective of the invention is to provide a system and device for real-time creditworthiness assessment of SMEs using autonomous, cooperative AI agents. A further objective is to establish an adaptive learning pipeline that continuously refines creditworthiness assessments based on transaction, behavioral, and operational signals. Another objective is to ensure the transparency of creditworthiness assessments through explainable AI models that satisfy both lenders and regulators. A further objective is to develop a hardware-implemented machine structure that can be deployed in banks, microfinance institutions, or data centers, enabling seamless integration of the assessment framework into existing financial infrastructure. Summary of the invention

[0017] The invention comprises a system with multiple autonomous AI agents, including data collection, evaluation, verification, and feedback agents. These are interconnected via secure APIs and orchestrated through an event-driven, real-time architecture. The system processes semi-structured and unstructured data from various sources, such as GST filings, bank APIs, e-invoices, inventory records, and sentiment data. It then generates a composite credit score that is continuously updated based on behavioral signals from SMEs.

[0018] The invention also provides a device configuration with a dedicated rating engine that integrates computing processors, secure memory modules, FPGA / GPU accelerators, blockchain-based integrity registers, and a user interface. The engine is structurally designed as a modular credit rating device that can be installed in financial institutions or deployed as a field unit for SME clusters. It enables localized, real-time credit rating while ensuring federated synchronization with central systems.

[0019] The present invention was developed with the aim of overcoming the inherent limitations of conventional and existing AI-based credit scoring systems for small and medium-sized enterprises (SMEs). A key objective of the invention is to provide an adaptive and intelligent real-time framework for SME credit scoring. This framework utilizes a network of autonomous software agents capable of independently collecting data, performing risk assessments, checking explainability, and conducting feedback-based learning. Another important objective of the invention is to ensure that the system responds to the dynamic operational circumstances of SMEs by continuously updating credit scores based on new transaction, behavioral, and environmental signals, thereby generating credit assessments that reflect the current state of the business.A further objective of the invention is to improve transparency and regulatory compliance in credit decisions by incorporating explainable artificial intelligence techniques that generate interpretable results and verifiable decision paths accessible to both lenders and regulators. Furthermore, the invention aims to integrate federated data collection functions from a variety of structured and unstructured sources, including, but not limited to, bank records, tax returns, digital invoices, social media data, and inventory management systems, ensuring that no data gap compromises the accuracy of credit assessments.Another objective is the development and provision of a device or machine implementation that physically operationalizes the system and includes dedicated computing modules, secure data processing hardware, blockchain-based verification components, and user-friendly interfaces. This will allow financial institutions to deploy the invention as a standalone device or as part of a distributed field infrastructure. A further objective of the invention is scalability and adaptability, enabling the system to be used not only by large banks but also by microfinance institutions and small regional lenders, thus ensuring equitable access to advanced credit scoring technology across the financial sector.The aim of the invention is to create a robust, understandable and real-time credit scoring solution for SMEs that promotes financial inclusion, reduces systemic risks and builds trust between businesses and credit institutions. BRIEF DESCRIPTION OF THE FIGURE

[0020] These and other features, aspects, and advantages of the present invention will be better understood if the following detailed description is read with reference to the accompanying drawing, in which the same symbols consistently represent the same parts. The following applies: Fig. Figure 1 shows a block diagram of the system and device for agent-based AI-based real-time creditworthiness assessment of small and medium-sized enterprises.

[0021] Experts will also recognize that the elements in the drawing are shown for the sake of simplicity and are not necessarily to scale. For example, the flowcharts illustrate the process by highlighting the main steps to enhance understanding of the aspects of this disclosure. Furthermore, with regard to the design of the device, one or more components of the device may be represented in the drawing by conventional symbols, and the drawing may show only the specific details relevant to understanding the embodiments of this disclosure, so as not to clutter the drawing with details that are readily apparent to those skilled in the art after reading this description. Detailed description of the invention

[0022] For a better understanding of the inventive principles, reference is made below to the embodiment shown in the drawing, which is described in specific terminology. However, this does not limit the scope of the invention. Changes and further modifications of the illustrated system, as well as further applications of the inventive principles, are possible, as would normally occur to a person skilled in the art in the field of invention.

[0023] It is clear to the person skilled in the art that the preceding general description and the following detailed description are exemplary and explanatory of the invention and are not intended as a limitation of it.

[0024] References in this specification to “an aspect”, “another aspect”, or similar expressions mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, occurrences of the expressions “in one embodiment”, “in another embodiment”, and similar expressions in this specification may all refer to the same embodiment, but need not.

[0025] The terms "includes," "include," or other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that includes a list of steps may not only contain those steps but may also include other steps not expressly listed or inherent in such process or method. Likewise, the statement "includes..." in the case of one or more devices, subsystems, elements, structures, or components does not, without further limitations, preclude the existence of other devices, subsystems, elements, structures, components, or additional devices, subsystems, elements, structures, or components.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by a person skilled in the art in the field of the invention. The system, methods, and examples provided here serve only for illustration and are not to be construed as a limitation.

[0027] Embodiments of the present disclosure are described in detail below with reference to the attached drawing.

[0028] In Fig.Figure 1 shows a block diagram of a system and device for agent-based, AI-based real-time credit scoring for small and medium-sized enterprises. The system 100 comprises: a data ingestion processing unit (102) configured to ingest and preprocess heterogeneous data streams from structured and unstructured sources, including financial transaction APIs, tax databases, e-invoicing registers, inventory management systems, social sentiment feeds, and IoT-enabled devices, with preprocessing including schema harmonization, anomaly detection, and encryption-based integrity preservation;an ensemble of scoring agents (104) configured to generate dynamic credit scores, the ensemble comprising machine learning models including gradient-enhanced decision trees, recurrent neural networks, and graph neural networks, the graph neural networks being configured to assess relational dependencies between SMEs, suppliers, and customer nodes in a business ecosystem; a verification agent (106) configured to apply explainable AI methods, including SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations), to generate regulatory-compliant reasoning paths for each credit score;a feedback agent (108) configured to update the ensemble of scoring control units using reinforcement learning signals derived from observed repayment behavior, payment default trends, and fraud detection markers; and a user interface (110) configured to display timestamped credit histories, confidence intervals, and explanatory justifications, the entire system running on a containerized cloud-native infrastructure with blockchain-based ledgers for version control of each score update.

[0029] In one embodiment, the data ingestion processing unit (102) uses a federated data access protocol configured to extract sensitive SME data records without centralizing raw data. The protocol also includes secure multi-party computations and differential data protection layers, so that lenders receive risk characteristics while the underlying transaction or tax details remain encrypted locally at the data source.

[0030] In one embodiment, the ensemble of the rating agent (104) manages two feature stores: The first is a long-term historical feature store that stores aggregated behavioral attributes such as average monthly cash inflows and seasonally adjusted sales growth, and the second is a short-term, event-driven store that stores momentary anomalies such as late invoice payments, sudden order cancellations, or inventory depletion. Both stores are dynamically synchronized and used together to update the creditworthiness of SMEs in near real time.

[0031] In one embodiment, the verification agent (106) generates human-readable compliance reports by mapping machine learning-derived features to regulatory categories. This mapping is performed by executing an ontology-driven semantic alignment that links input variables such as tax compliance or inventory turnover with standardized risk factors defined by financial supervisory authorities. The compliance reports are digitally signed and appended to a blockchain ledger to ensure immutability and traceability.

[0032] In one embodiment, the feedback processing unit (108) is trained using reinforcement learning, which includes a policy gradient technique, wherein the policy is updated when the repayment or default results deviate from the predicted credit scores, and wherein the agent adjusts the hyperparameters of the scoring ensemble to reduce the cumulative prediction error over successive credit cycles, thereby providing adaptive recalibration of the model in response to the evolving behavior of SMEs.

[0033] In one embodiment, the user interface (110) is executed on a distributed edge cloud interface. The user interface includes an interactive visualization layer that supports drill-down analysis of SME-specific indicators, a risk heatmap created using dynamic graph embeddings, and a temporal anomaly tracker configured to notify financial officials in real time when creditworthiness deviations exceed a predefined confidence threshold. The user interface also supports regulatory audit access for multiple users with different levels of authorization.

[0034] In one implementation, each credit rating update is cryptographically signed using asymmetric key pairs generated in a Trusted Platform Module (TPM). The TPM is embedded in the system infrastructure, and the signatures are anchored in a distributed blockchain ledger, enabling lenders and regulators to verify credit history and prevent retroactive manipulation of SME credit documentation.

[0035] In one embodiment, the containerized cloud-native infrastructure is orchestrated by Kubernetes clusters that include load balancing mechanisms to scale the scoring control unit ensemble in response to transaction spikes, using message queues such as Apache Kafka for asynchronous data streaming and forwarding latency-sensitive inference requests to GPU-accelerated nodes, while audit and compliance tasks are assigned to CPU-optimized nodes to optimize computing power.

[0036] In one embodiment, the graph neural network implemented in the scoring control unit ensemble encodes SME units as nodes and transaction dependencies as weighted edges, with temporal decay factors assigned to the edges so that older relationships have a diminishing influence on the creditworthiness check, and with the network additionally being trained with edge-level fraud detection signals to identify anomalous relationships that may indicate collusion or circular trading.

[0037] In one embodiment further comprising a device configuration that is configured as a standalone credit rating device, wherein the device comprises: a multi-core processor array with integrated FPGA and GPU accelerators, optimized for low-latency AI inference; a secure storage module that is divided into volatile and non-volatile layers, with the volatile layer handling volatile SME data streams and the non-volatile layer storing encrypted feature embeddings; an integrity register embedded in the blockchain, implemented in a hardware-trusted platform module to store immutable versions of credit score outputs; an interactive user interface integrated into the device housing that allows financial officials to directly monitor the risk profiles of SMEs; and A tamper-proof enclosure with biometric access control and redundant power supply to ensure secure, uninterrupted operation in institutional or remote environments.

[0038] Each agent operates independently but communicates securely with others via standardized APIs and event-driven orchestration frameworks. The agents are deployed in a containerized, cloud-native infrastructure that ensures scalability, fault tolerance, and low latency. By combining ensemble machine learning models, graph neural networks, explainable AI modules, and blockchain-based ledgers, the system offers a comprehensive solution that not only delivers accurate credit scores but also guarantees transparency, compliance, and adaptability.

[0039] The credit scoring process begins with the data collection and processing unit (DCU), which connects to multiple heterogeneous data sources. These sources include structured data such as bank APIs, tax databases, and government GST registers, as well as semi-structured or unstructured data such as e-invoices, POS transaction logs, inventory management systems, and even IoT sensors tracking supply chain movements. Social sentiment feeds and digital reputation signals can also be integrated, particularly in cases where SMEs have limited formal financial documentation. The DCU applies schema harmonization techniques to classify diverse formats into standardized functional sets and uses anomaly detection techniques to filter out corrupted or fraudulent entries.To protect data privacy, federated learning protocols are used, where sensitive raw data remains at the source, while only anonymized features are transmitted to the central system. Encryption mechanisms, including secure multi-party computation and differential data protection levels, ensure that financial institutions can gain insights without compromising confidentiality.

[0040] The system (100) is enabled by concrete hardware and tightly integrated software-accelerated components that realize each of the aforementioned elements in physical form: The multitude of autonomous agents is embodied as distributed computing nodes (edge ​​servers, virtual machines or dedicated devices), each equipped with network interface controllers, secure element modules and hardware cryptography accelerators to enforce secure communication protocols and mutual authentication;The data ingestion processing unit is implemented as a bank of NICs, API gateway devices, message brokers, and I / O controllers that feed into preprocessing blades containing CPUs, GPUs / TPUs, FPGAs, or dedicated ASICs, where schema harmonization logic, anomaly detection filters, and encryption / integrity engines (including hardware HMAC and AES coprocessors) operate with streamed structured and unstructured inputs, with local SSD caches and high-throughput storage arrays for deployment;The scoring control unit ensemble is implemented on clustered model-serving hardware with GPU / TPU racks and graph accelerator cards hosting gradient-enhanced decision tree inference machines, recurrent neural network instances, and graph neural network cores – interconnected via a low-latency fabric and orchestrated across model-serving containers. This allows relational dependencies between SMEs, suppliers, and customers to be computed in memory across adjacency stores and fast NVMe-supported graph databases. The audit controller runs in secure enclaves or trusted execution environments on server platforms and is coupled with explainability computation modules (CPU / GPU tasks executing SHAP and LIME routines) that write immutable, timestamped justification paths to tamper-proof, append-only memory or a write-ahead ledger.The feedback processing unit is implemented in the form of telemetry collectors, streaming analytics nodes, and reinforcement learning accelerators that subscribe to event streams via Kafka-style brokers, perform reward signal calculations on GPU / TPU hardware, and forward policy updates to the scoring ensemble via secure CI / CD pipelines; the user interface is physically implemented as load-balanced web and API servers with GPU-capable rendering backends, time-stamped SQL / NoSQL stores for credit histories and confidence interval data, and hardware-enforced authentication modules for role-based access;and the entire deployment is implemented on a containerized, cloud-native infrastructure running on physical hosts or bare-metal clusters with orchestration controllers (e.g., Kubernetes control plane nodes), hardware security modules (HSMs) for key management, and a set of blockchain validation nodes or ledger appliances that provide version-controlled, cryptographically anchored registers for each point update – these concrete combinations of processors, accelerators, memory, network hardware, and secure modules ensure the practical activation and concrete implementation of the described system.

[0041] Once the data is harmonized, it is processed in dual feature stores that support both long-term and short-term perspectives. The long-term feature store aggregates historical behavioral characteristics such as average monthly cash flows, multi-year repayment patterns, and seasonally adjusted sales growth curves. The short-term, event-driven feature store, on the other hand, captures real-time anomalies such as sudden payment defaults, unusual order cancellations, or rapid inventory depletion. Scoring agents continuously access both stores to ensure a balanced view that responds to immediate operational changes as well as reflects longer-term business behavior.

[0042] The Scoring Control Unit ensemble forms the computing core of the invention. It combines gradient-based decision trees such as XGBoost and LightGBM for structured numerical data, recurrent neural networks for sequential transaction histories, and graph neural networks for the relational analysis of SME networks. The graph neural network component represents SMEs as nodes and transaction dependencies as edges, applying temporal decay factors so that older transactions have a diminishing impact on creditworthiness. This allows the system to dynamically model the development of SME networks and identify risks that can propagate through supplier or customer relationships. Furthermore, fraud detection signals can be embedded in edge attributes, enabling the network to detect suspicious patterns such as circular trading, invoice inflation, or collusion between multiple SMEs.

[0043] The scoring results are supplemented with explanatory layers to ensure regulatory compliance and strengthen stakeholder confidence. The audit controller uses SHAP values ​​to quantify the contribution of individual characteristics to each credit score and LIME to generate localized model explanations for specific SMEs. These results are further mapped to regulatory categories using an ontology-based semantic alignment. For example, a raw characteristic such as inventory turnover is mapped to the standardized risk factor of operational efficiency. This mapping generates compliance-ready reports that can be presented to regulators and lenders in an easily understandable format. Each report is digitally signed and stored in a blockchain-based integrity register to ensure that historical justifications cannot be retrospectively altered.

[0044] The feedback processing unit closes the loop through continuous results monitoring. When a loan is granted to an SME, the agent observes repayment behavior, payment delays, and fraud. These observations are transformed into reinforcement learning signals, which are fed back to the scoring control unit's ensemble. A policy gradient technique adjusts the ensemble's hyperparameters, weights, and feature priorities based on deviations between predicted and actual results. Across successive loan cycles, the feedback processing unit reduces the cumulative prediction error and effectively calibrates the system to adapt to changing economic conditions and SME behavior. This function ensures that the system remains accurate even in volatile or data-poor environments.

[0045] All these processes are orchestrated in a containerized, cloud-native environment. Kubernetes clusters manage the elastic scaling of agents and dynamically allocate compute resources based on transaction volume. Low-latency inference queries are routed to GPU-accelerated nodes, while compliance checks and audit functions are assigned to CPU-optimized nodes, thus optimizing resource utilization. Apache Kafka or similar message queues handle the asynchronous streaming of events, ensuring that data ingestion, scoring, auditing, and feedback occur in parallel and without bottlenecks. Each credit score update is timestamped, cryptographically signed with asymmetric key pairs generated in a trusted platform module, and immutably stored in a blockchain ledger.This provides regulators and lenders with verifiable and unmanipulated credit histories, thereby strengthening trust and compliance.

[0046] For user interaction, the system offers a user interface that visualizes the credit histories of SMEs. The interface supports interactive drill-down into SME-specific indicators, displays risk heatmaps from dynamic graph embeddings, and issues anomaly alerts when credit scores deviate beyond predefined confidence intervals. Access to the interface is tiered according to user rights: loan officers can view detailed SME reports, while regulators can access compliance logs and historical justifications. The interface operates via distributed edge cloud interfaces, allowing local finance staff to use it in regions with intermittent connectivity and synchronize with central servers when bandwidth is available.

[0047] The invention also discloses a physical device implementation of the system. The device can be installed as a standalone credit scoring device in banks or financial institutions and comprises a multi-core processor with integrated FPGA and GPU accelerators for performing low-latency AI inferences. The device also includes a secure memory module divided into volatile and non-volatile layers. Transient SME data is processed in volatile memory, and only encrypted embeddings are stored permanently. A blockchain-anchored integrity register is implemented in a hardware-based trusted platform module embedded in the device, ensuring that score updates remain tamper-proof. The device housing features an embedded user interface that allows financial staff to directly monitor SME risk profiles without requiring external hardware.To protect the device, the housing is equipped with tamper-proof functions, biometric access controls and redundant power supplies to ensure uninterrupted operation.

[0048] In alternative embodiments, the device can be configured as a portable credit scoring terminal for use in remote or underserved regions. Such a terminal could be rugged, equipped with satellite communication modules for off-grid operation, and solar-powered. The terminal would maintain offline-to-online synchronization, allowing credit assessments to be performed locally even without an internet connection. Later, once a network connection is available, the records would be synchronized with the central blockchain registry. This embodiment addresses the needs of rural SME clusters, microfinance institutions, and developing countries where reliable connectivity cannot be guaranteed.

[0049] By combining agent-based AI software architecture and dedicated device execution, the invention offers a robust, adaptive, and transparent system for real-time creditworthiness assessment of SMEs. It addresses the shortcomings of existing systems through the integration of multi-source data fusion, graph-based contextual reasoning, explainable decision-making, feedback loops for reinforcement learning, and tamper-proof audit trails. This enables financial institutions of all sizes to fairly and effectively assess the creditworthiness of SMEs in dynamic environments.

[0050] In one embodiment, the system uses a multi-agent architecture. Data agents establish connections to external APIs such as tax registers, payment gateways, and enterprise resource planning platforms, and extract structured and unstructured datasets. These datasets are then processed into feature vectors that are continuously updated in long-term and short-term feature stores.

[0051] Scoring agents utilize these functions with the help of ensemble machine learning models such as XGBoost, LightGBM, and graph neural networks. By employing graph neural networks, the system can capture dependencies between SME clusters, suppliers, and customers, thus performing a relational credit risk assessment. The scoring agents generate risk forecasts and decision rationales, which are further validated by audit agents. Audit agents implement explainability techniques such as SHAP and LIME to generate human-interpretable rationales, ensuring that credit decisions comply with regulatory requirements.

[0052] Feedback agents enhance the learning process by observing post-loan signals, such as repayment behavior, payment default trends, and fraud indicators. Reinforcement learning loops ensure that assessment guidelines are dynamically recalibrated with each interaction, improving resilience against data drift.

[0053] The system runs on a cloud-native container infrastructure and utilizes orchestration platforms such as Kubernetes for flexible scaling, Kafka for real-time streaming, and blockchain-based ledgers for version control and data integrity. Each score update is digitally signed and timestamped, creating a verifiable audit trail. A user interface displays credit score histories, risk alerts, and confidence intervals.

[0054] The machine's physical structure is designed as a rack or kiosk version and features a tamper-proof enclosure, biometric access for authorized personnel, and a redundant power supply for uninterrupted operation. The machine ensures local inference capability even in environments with limited internet connectivity and synchronizes with a central cloud cluster as soon as the connection is restored.

[0055] In an alternative embodiment, the device can be configured as a portable credit scoring terminal, resembling a rugged laptop or edge device, for deployment in remote SME centers. This configuration includes satellite communication modules, offline memory synchronization, and solar-powered operation for field deployment in regions with limited access to banks.

[0056] The present invention relates to the fields of artificial intelligence, financial analysis, and credit risk assessment. More specifically, it relates to the development of an agent-based AI framework and an associated device for the real-time assessment of the creditworthiness of SMEs. The invention is based on the combination of machine learning, explainable AI, distributed multi-agent coordination, federated data integration, and blockchain-based audit systems. By providing a scalable and adaptive architecture that continuously processes data streams from multiple sources, generates explainable credit decisions, and updates models through feedback loops, the invention addresses long-standing challenges in SME financing.It ensures that financial institutions can carry out transparent, regulatory-compliant and dynamic credit assessments tailored to the operational circumstances of SMEs in different economic contexts.

[0057] The drawing and the preceding description show examples of embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another embodiment. For example, the sequence of the processes described here can be changed and is not limited to the manner described here. Furthermore, the actions of a flowchart need not be implemented in the sequence shown; nor does it necessarily have to be performed by all actions. Actions that are not dependent on other actions can also be performed in parallel with the other actions. The scope of the embodiments is in no way limited by these specific examples.Numerous variations are possible, whether explicitly stated in the specification or not, such as differences in structure, dimensions, and material usage. The range of embodiments is at least as broad as specified in the following claims.

[0058] Advantages, further benefits, and solutions to problems have been described above with regard to specific embodiments. However, the advantages, benefits, solutions to problems, and all components that may lead to a particular advantage or solution occurring or becoming more apparent are not to be construed as critical, necessary, or essential features or components of any or all claims. REFERENCES 100 A system and device for agent-based AI-supported real-time credit rating of small and medium-sized enterprises. 102 Data acquisition and processing unit 104 Gate counter 106 Audit Officer 108 Feedback processing unit 110 User interface QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 20200236238A1 [0003, 0008] US 20210143599A1 [0003, 0009] US 20200341765A1 [0003, 0010] US 20190082316A1

[0007] US 20210120809A1

[0011]

Claims

[1] A system for real-time creditworthiness assessment of small and medium-sized enterprises (SMEs) based on agent-based artificial intelligence (AI), the system includes: a large number of autonomous agents connected to each other via secure communication protocols; a data ingestion processing unit configured to capture and preprocess heterogeneous data streams from structured and unstructured sources, including APIs for financial transactions, tax databases, e-invoicing registers, inventory management systems, social sentiment feeds, and IoT-enabled devices, with preprocessing including schema harmonization, anomaly detection, and encryption-based integrity preservation; an ensemble of scoring control units configured to generate dynamic credit scores, wherein the ensemble includes machine learning models, including gradient-enhanced decision trees, recurrent neural networks, and graph neural networks, the graph neural networks being configured to assess relational dependencies between SMEs, suppliers, and customer nodes in a business ecosystem; an audit controller configured to generate regulatory-compliant justification paths according to each credit score; a feedback processing unit configured to update the evaluation control unit ensemble using reinforcement learning signals derived from observed repayment behavior, payment default trends, and fraud detection markers; and a user interface configured to display timestamped credit histories, confidence intervals and explanatory justifications, with the entire system running on a containerized cloud-native infrastructure with blockchain-based ledgers for version control of each point update. [2] System according to claim 1, wherein the data ingestion processing unit uses a federated data access protocol configured to extract confidential SME data sets without centralizing raw data, the protocol further comprising secure multi-party computations and differential data protection levels so that lenders receive risk characteristics while the underlying transaction or tax details remain encrypted locally at the data source. [3] System according to claim 1, wherein the scoring control unit ensemble manages two feature stores, the first of which is a long-term historical feature store that stores aggregated behavioral attributes, such as average monthly cash inflows and seasonally adjusted sales growth, and the second of which is a short-term, event-driven store that stores momentary anomalies, including late invoice payments, sudden order cancellations or inventory depletion, wherein both stores are dynamically synchronized and used together to update the credit ratings of SMEs in near real time. [4] System according to claim 1, wherein the audit controller generates human-readable compliance reports by assigning machine learning-derived features to regulatory categories, the assignment being achieved by performing ontology-driven semantic alignment that links input variables such as tax compliance or inventory turnover with standardized risk factors defined by financial supervisory authorities, and wherein the compliance reports are digitally signed and attached to a blockchain register to ensure immutability and traceability. [5] System according to claim 1, wherein the feedback processing unit is trained using reinforcement learning, which includes a policy gradient technique, wherein the policy is updated when the repayment or default results differ from the predicted credit scores, and wherein the agent adjusts the hyperparameters of the scoring ensemble to reduce the cumulative prediction error over successive credit cycles. [6] System according to claim 1, wherein the user interface is executed on a distributed edge cloud interface and comprises an interactive visualization layer that supports drill-down analysis of SME-specific indicators, a risk heatmap generated using dynamic graph embeddings, and a temporal anomaly tracker configured to notify financial officials in real time when creditworthiness deviations exceed a predefined confidence threshold, and wherein the user interface also supports regulatory audit access by multiple users with different levels of authorization. [7] System according to claim 1, wherein each credit rating update is cryptographically signed using asymmetric key pairs generated in a Trusted Platform Module (TPM), the TPM being embedded in the system infrastructure, and wherein the signatures are anchored in a distributed blockchain ledger so that lenders and regulators can verify historical credit histories and prevent retrospective manipulation of SME credit documentation. [8] System according to claim 1, wherein the containerized cloud-native infrastructure is orchestrated by Kubernetes clusters which include load balancing mechanisms to scale the scoring control unit ensemble in response to transaction spikes, wherein message queues such as Apache Kafka are used for asynchronous data streaming and wherein latency-sensitive inference requests are forwarded to GPU-accelerated nodes, while audit and compliance tasks are assigned to CPU-optimized nodes to optimize computing power. [9] System according to claim 1, wherein the neural graph network implemented in the ensemble of the rating control unit encodes SME units as nodes and transaction dependencies as weighted edges, wherein temporal decay factors are assigned to the edges so that older relationships have a diminishing influence on the creditworthiness check, and wherein the network is also trained with edge-level fraud detection signals to identify anomalous relationships that indicate collusion or circular trading. [10] System according to claim 1, further comprising a device embodiment configured as a standalone credit rating device, wherein the device comprises: a multi-core processor array with integrated FPGA and GPU accelerators, optimized for low-latency AI inference; a secure storage module that is divided into volatile and non-volatile layers, with the volatile layer handling volatile SME data streams and the non-volatile layer storing encrypted feature embeddings; an integrity register embedded in the blockchain, implemented in a hardware-trusted platform module to store immutable versions of credit score outputs; an interactive user interface unit that is integrated into the device housing and allows financial officials to directly monitor the risk profiles of SMEs; A tamper-proof enclosure with biometric access control and redundant power supply to ensure secure, uninterrupted operation in institutional or remote environments.

Images