Code review method using incremental and artificial intelligence-assisted code processing with multiple virtual reviewers

Description

[001] This source code review method is applicable in software development processes that use version control and code review. Code is obtained from a repository, processed, and then Artificial Intelligence (or AI) is used to analyze this code by multiple virtual reviewers. This review is then made available to the user, showing the identified problems and their solutions. The method can be integrated into CI / CD pipelines (continuously integrated / continuously deployable source code), automating the detection of problems in modified code. It is useful for both corporate teams and open source projects, ensuring compliance with quality standards and good programming practices. Furthermore, it can be used in software development training, providing automated and consistent feedback.This method is ideal for optimizing the time and efficiency of reviews, ensuring quality and security in the code. FUNDAMENTALS OF THE INVENTION

[002] Currently, automated code review tools are widely available and used to ensure software quality. Tools such as SonarQube, Checkstyle, and PMD are examples of systems that analyze source code for style violations, excessive complexity, or security vulnerabilities. These tools rely on predefined rules and static lists of known issues, being highly specialized and limited by the coverage of their databases.

[003] In addition, Artificial Intelligence (AI)-based solutions have begun to emerge in the field of code review. Some approaches use natural language models, such as Codex and GitHub Copilot, to suggest improvements or detect problems in the code during development. However, these solutions present results that lack precision and consistency when using a single prompt or Petition 870250007435, dated 01 / 29 / 2025, page 6 / 51 2 / 15 require multiple iterations to review and correct code, which can increase computational cost. One of the most common methods used by AI-assisted review systems is called Evaluation Prompting, in which the AI ​​reviews its own outputs from a second iteration. While this method improves the consistency of reviews, it is computationally expensive, multiplying the number of prompts.

[004] Version control and collaborative review-based solutions, such as GitHub Pull Request Reviews and Gerrit, allow human developers to manually review incremental code changes without AI automation, requiring considerable human effort, which often leads teams to not adopt the practice of code review.

[005] Although AI language models, such as general-purpose ones (e.g., GPT), can be adapted for code review tasks, current solutions often lack accuracy and consistency or rely excessively on specialized training or intensive use of chained prompts.

[006] As some specific problems of the state of the art, we can cite the following points: • Lack of Consistency in AI-Assisted Reviews: AI-based solutions, such as Codex or GitHub Copilot, often provide inconsistent or unpredictable results, especially when applied to multiple code contexts without specialized training; • High Computational Cost of Evaluation Prompting: The use of evaluation, where the AI ​​reviews its own outputs on multiple prompts, significantly increases the computational cost; • Dependence on Static Rules: Tools like SonarQube and Checkstyle rely on predefined rules, which may not cover all scenarios or may not adapt to specific projects; • Difficulty in Identifying Problems in an Incremental Context: Formats such as unidiff, used in version control systems, have limitations in providing sufficient context for AI to consistently evaluate changes; Petition 870250007435, dated 01 / 29 / 2025, page 7 / 51 3 / 15 • Inefficiency in Locating Problems: Automated review tools often generate separate reports or listings that do not clearly identify the exact lines where problems occur; • Difficulty in Proposing Complete Solutions: Many code review tools point out problems, but do not offer clear solutions or explanations; • Need for intensive training of specialized models. Currently, to achieve accuracy in automated code reviews using AI, many solutions rely on highly trained language models for specific use cases, such as in areas of security, efficiency, or good coding practices. This training process demands significant computational effort, as well as time and resources to develop and maintain models tailored to different programming languages ​​or application contexts. STATE OF THE ART

[007] US patent 20210089992A1 [1] describes a method for recommending code reviewers in an automated way, based on several factors such as expertise, past contributions, and availability. This patent describes a patented method that aims to automate the recommendation of human reviewers for submitted code, rather than the code review itself. It uses a learning phase in which an artificial intelligence agent understands the contextual structure of code regions, mapping them to a distributed representation. In the recommendation phase, the agent generates a ranked list of recommended reviewers for any submitted code review request, presenting this list to a visualization device, thus optimizing selection and increasing the efficiency of the code review process.

[008] US patent 20210182031A1 [2] describes a patented system that examines software source code against reference source code for patterns that may indicate the presence of bugs. This is done automatically, without the need for constant manual intervention. Using advanced techniques such as machine learning and static analysis algorithms, the system identifies patterns and anomalies in the code that are common in software bugs. Once a potential bug is identified, the system generates detailed reports. These reports include. Petition 870250007435, dated 01 / 29 / 2025, page 8 / 51 4 / 15 Information about the location of the bug in the code, the nature of the problem, and possibly suggestions on how to fix the error.

[009] The solution described in the article AI-Assisted Assessment of Coding Practices in Modern Code Review [1] involves the development and implementation of AutoCommenter is a system based on a large language model (LLM) that automatically learns and applies best coding practices during code review. The goal of AutoCommenter is to automate the detection of best practice violations, providing timely feedback to code authors and allowing human reviewers to focus on the overall functionality of the code. Using a T5-based text-to-text transformation model, the system is trained to identify best practice violations in C++, Java, Python, and Go, receiving a code snippet as input and generating the location of the violation and the URL of the corresponding best practice document.Training the model involves creating examples from real code review data, where human comments with best practice URLs are extracted and used to train the model, going through preprocessing, dataset curation, and fine-tuning. The AutoCommenter inference infrastructure includes a central service that analyzes code files, builds the model input, queries the model, and filters out low-quality predictions before returning the remaining predictions.

[010] The article “Resolving Code Review Comments with Machine Learning” [4] describes an ML assistance feature to reduce the time spent resolving code review comments on Google. The text describes the creation of a machine learning (ML)-based software engineering assistant to improve the speed of code changes, focusing on the automated resolution of code review comments. The assistant suggests edits to resolve specific comments, which can be viewed and applied by the authors. The problem modeling is done as a text-to-text task using a T5-based Transformer architecture, with a training corpus of over 3 billion examples. The model is tuned to maximize the accuracy of the predictions, which are accompanied by a confidence probability. The practical implementation Petition 870250007435, dated 01 / 29 / 2025, page 9 / 51 5 / 15 involves listening to review feedback, generating inputs for the model, predicting code edits, and exposing those edits to users through integrated systems, collecting interaction logs to generate insights.

[011] Amazon CodeGuru Reviewer [5] is a service that integrates with some DevOps environments and uses program analysis and machine learning to detect potential defects that are difficult for developers to find and offers suggestions to improve your Java and Python code. It does not point out source code errors exactly where they occur, only making general comments per file.

[012] Swimm [6] is a generative AI-powered solution in the form of an IDE extension (“plugin”) that assists developers in creating source code documentation.

[013] AI Code Review Action [7] is a generative AI-powered code review solution, integrated exclusively into the Github DevOps environment, which provides suggestions for corrections.

[014] Codacy [8] is a generative AI-backed code review solution, integrated into the DevOps environment, with the primary purpose of collecting source code metrics.

[015] DeepCode AI [9] is a generative AI-powered code review solution, integrated into the DevOps environment, that points out security issues.

[016] Code Climate

[10] is an engineering metrics platform that makes use of generative AI-powered code review, integrated into the DevOps environment.

[017] Code Rabbit

[11] is a generative AI-powered code review solution, integrated exclusively into the Github DevOps environment, which provides suggestions for corrections.

[018] Thus, given what is observed in the state of the art presented, there is room for innovations that combine generative Artificial Intelligence, collaborative review methodologies, and efficient processing of incremental changes in code, as proposed by the method shown here, which improves the consistency of reviews without significantly increasing the computational cost, which is what is intended to be achieved in the patent application shown here. Petition 870250007435, dated 01 / 29 / 2025, page 10 / 51 6 / 15 OBJECTIVES AND ADVANTAGES OF THE INVENTION

[019] The main objective of this patent application is to provide an innovative and efficient solution for AI-assisted code review, solving problems of inconsistency, high computational cost, and low adaptability present in current solutions. Specifically, the invention seeks the following: • Increase the consistency of code reviews: Through the methodology of multiple virtual reviewers, the tool aims to improve the accuracy of code problem identification, overcoming the unpredictability of existing AI approaches that use a single prompt (a single processing request for the AI, per source code file); • Reduce computational cost: By using a single prompt to reach consensus among virtual reviewers, the method shown here reduces computational cost compared to techniques such as Evaluation, which require multiple AI interactions to verify its own consistency; • Avoiding the need for specialized training: The system is designed to work efficiently with general-purpose language models, eliminating the need to train or adjust the AI ​​for specific projects or languages, as required in other solutions on the market; • Improve accuracy in identifying incremental problems: The invention processes source code in unidiff format in an innovative way, applying changes to the original code and ensuring that the AI ​​receives sufficient context to identify and report problems in the correct lines; • Provide more detailed and actionable feedback: The tool offers clear descriptions of problems, explanations of the impact, and suggestions for correction, making it easier for developers to resolve issues; • Automate the recording of issues directly in the code: The method shown here records issue notes in the exact lines of code, unlike solutions that generate separate reports or lists far removed from the problem. Petition 870250007435, dated 01 / 29 / 2025, page 11 / 51 7 / 15

[020] These advantages make the invention an advanced and more efficient solution for code review, overcoming the challenges and limitations of current solutions on the market. GENERAL DESCRIPTION OF THE INVENTION

[021] The method shown here is an algorithm created to interact programmatically with the APIs of a manual source code (software) review tool and a generative artificial intelligence (AI), in order to obtain the proposed source code from the former and obtain from the latter the identification of problems in this code. The identified problems are recorded via the review tool's API to be displayed to the developer who will correct them.

[022] The method offers an inventive approach to AI-assisted code review, which uses the prompt technique known as tree of thoughts, implementing a methodology of steps and virtual review agents with a single prompt. Initially, the virtual reviewers examine the source code and identify problems, and then they examine the problems found by the group of reviewers and carry out a collaborative vote, expressing agreement or disagreement individually.

[023] With the use of the multiple virtual reviewers methodology, the virtual collaborative voting step leads the artificial intelligence to a self-assessment of what it itself generated before pointing out problems.The method programmatically discards problems that have not obtained an arbitrary minimum number of votes, and therefore the result is significantly more consistent than that obtained with trivial instructions, even with a general-purpose language model, and this has two important implications: • Absence of the need to employ human and computational effort in specializing the language model to transform it into a specific-purpose model. • The absence of a need to chain prompts to achieve consistency, which would multiply the effort and computational cost of obtaining each response. Petition 870250007435, dated 01 / 29 / 2025, page 12 / 51 8 / 15

[024] The method soon implements an innovative communication protocol with AI, with the following differentiating characteristics when this method is compared with the state of the art: • Processing of source code received in “unidiff” format, preparing it for unambiguous and consistent identification of the problematic code location. • Invoking the AI ​​with meta-instruction messages at the prompt that instruct it to generate a structured output format that takes into account the preparation of the source code; • Processing of the AI ​​response, which follows the pattern of the previous format.

[025] The result is the ability to programmatically record comments alongside the code review tool with identification of the corresponding section (start and end, each marked by a pair of “line” and “right-hand characters”). DESCRIPTION OF THE FIGURES

[026] The invention will be described in detail below, and for better understanding, reference will be made to the attached drawings, in which are represented: • Figure 1: Diagram showing the components used to implement the method; • Figure 2: Flowchart showing the main steps of the method; • Figure 3: Flowchart showing Step 1 of the method with all its substeps in its preferred implementation; and • Figure 4: Flowchart showing Step 3 of the method with all its substeps in its preferred implementation, and showing the use of virtual reviewers and collaborative voting on each problem and solution result generated by the reviewers. DETAILED DESCRIPTION OF THE INVENTION

[027] The CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS shown here is an algorithm capable of interacting programmatically with the API of a manual source code review tool (F) Petition 870250007435, dated 01 / 29 / 2025, page 13 / 51 9 / 15 possessing a source code repository (R), and using generative Artificial Intelligence (AI), in order to obtain the proposed source code from the repository (R) and perform problem verification of this source code. The identified problems are recorded through the manual review tool's API (F) to be displayed, so that a developer can correct them. The method itself is hosted and executed on a computer (C), this computer being the user's local machine or even virtualized in the cloud. Specific implementation examples include the manual source code review tool (F) being the Azure DevOps cloud development environment, using the tool's repository (R) being Azure Repos, and the generative Artificial Intelligence (AI) being a GPT-40 instance. The method is performed through the sequential execution of four steps.

[028] The method begins with a Step 1 (1), in which the computer (C) obtains source code files from the repository (R), and performs the processing of the files, in order to establish a protocol for identifying the exact lines of code where each problem is located.

[029] The source code is obtained, in its main embodiment, in “unidiff” format, which allows incremental identification of changes in the source code. This format is unsuitable for sending to Artificial Intelligence (AI) for two reasons. The first is the absence of the rest of the file, since Artificial Intelligence (AI) needs a minimum of context to be able to reason consistently. The second reason is the difficulty in guiding Artificial Intelligence (AI), via prompt, to unequivocally identify the lines of source code in generating its return to the method.

[030] Thus, this Step 1 (1) is carried out, in specific and main implementation of the method, the following substeps: • Substep 1 (1S1): Obtaining by the computer (C) the version prior to the current one of the source code and archiving in unidiff format the source code change of selected files from the repository (R); • Substep 2 (1S2): Application by computer (C) of the changes to the unidiff source code of the files, obtaining a preview of the file after the change; Petition 870250007435, dated 01 / 29 / 2025, p. 14 / 51 10 / 15 • Substep 3 (1S3): Processing with sequential addition of source code lines to the files according to the unidiff.

[031] To obtain each file to be submitted to automatic review by Artificial Intelligence (AI), in a specific implementation using Azure Repos (R) repository, for example, the following requests are made to the Azure Repos (R) repository API to obtain the source code files: • Obtaining the comments made in the Pull Request; • Recording of commentary situations; • Obtaining the most recent iteration of the Pull Request, where iterations are the succession of versions of the same Pull Request; • Obtaining the files changed in a specific iteration of the Pull Request; • Obtaining general information about the Pull Request; • Adding comments per line of source code, per file, and per Pull Request; • Obtaining the change data for each file in unidiff format; • Obtaining the entirety of a given source code file.

[032] In Step 2 (2), the computer (C) constructs a prompt, written in natural language and then sent to the Artificial Intelligence (AI) along with the source code to be analyzed, for the use of said Artificial Intelligence (AI) for source code analysis, this prompt containing the following information: • Instruction message, which describes the task and methodology of virtual reviewers (VR) and voting (V), and includes the source code; • Meta-instruction message, which describes the information exchange protocol with Artificial Intelligence (AI); and • Randomness configuration parameters and response size parameter, adjustable at program runtime, in order to calibrate the Artificial Intelligence (AI) return for high accuracy and predictability.

[033] Specifically regarding the prompt instruction message, the following information is included therein: • Description of the task to be performed, that is, the analysis and search for problems in source code; Petition 870250007435, dated 01 / 29 / 2025, page 15 / 51 11 / 15 • Total number of virtual reviewers (VR); • Indication of the recurring stages and cycles of individual review and collaborative voting (V) on the validity of each problem and the impact attributed; • Programming language corresponding to the source code; and • Source code processed with line information, obtained in Step 1 (1).

[034] This prompt is thus used for the operation of Artificial Intelligence (AI).

[035] Step 3 (3) of the method uses the prompt created in Step 2 (2) for Artificial Intelligence (AI) operation and sending the source code to be analyzed, the Artificial Intelligence (AI) detecting problems in the source code and generating solutions to be suggested for each problem verified, this is done using multiple virtual review agents (VR) of the code and collaborative voting (V) for analysis of each problem detected in the source code, and returning the answer to the computer (C), this using the prompt technique known as tree of thoughts. Thus, a method is implemented using multiple virtual review agents (VR) using a single prompt.

[036] In normal Artificial Intelligence (AI) usage mode, i.e., without the indication of multiple reviewers (VR), the natural variability of AI response is subject to occasional errors, which are known as “hallucinations”. If the same prompt is executed repeatedly for a relatively large source code, with say at least a hundred lines, the chances are that the response will vary.

[037] With the multiple virtual reviewers (VR) methodology, each reviewer (VR) will naturally have random variability, but the collaborative voting (V) stage takes place internally in the Artificial Intelligence (AI) before the return and gives it the opportunity to disagree with its own inferences, that is, to disagree with the error pointed out by a virtual reviewer (VR) that may have been hallucinating.

[038] Although generative Artificial Intelligence (AI) is subject to hallucinations in text generation, it is significantly more effective at classification, that is, at judging the correctness of a generated response.

[039] The main way of carrying out Stage 3 (3) is through the specific implementation of the following sub-stages: Petition 870250007435, dated 01 / 29 / 2025, page 16 / 51 12 / 15 • Substage 1 (3S1): Virtual reviewers (VR), individually, examine the source code in the files, identify problems, and propose a solution for each problem pointed out in the source code; • Substage 2 (3S2): The reviewers (RV) examine each problem-solution pair indicated by the first reviewer (RV) and perform a collaborative vote (V), expressing agreement or disagreement individually on each problem-solution pair, this being repeated for the second reviewer (RV) and so on until the last reviewer (RV); • Substage 3 (3S3): The Artificial Intelligence (AI) returns the list of problem and solution pairs from each reviewer (RV) to the computer (C); and • Substage 4 (3S4): After the return from the Artificial Intelligence (AI), the computer (C) discards the problem and solution pairs that have had at least one disagreement, this substage thus eliminating the problem and solution pairs that have a high probability of being hallucinations.

[040] In the last Step 4 (4), the computer (C) processes the return from the Artificial Intelligence (AI) and it records comments using the review tool API (F) that indicate the problems pointed out line by line based on the processed file sent by the Artificial Intelligence (AI) in Step 3 (3) along with the indicated solution.

[041] One of the most important inventive features of the method is the ability to provide Artificial Intelligence (AI) with a structured output format through the prompt for pointing out problems. In conjunction with code processing to indicate lines with problems, the result is the production of a response that unequivocally and consistently identifies the position of each code problem.

[042] In the meta-instruction of the prompt constructed in Step 2 (2), as mentioned, the output portion of the protocol is described to Artificial Intelligence (AI), establishing the structured pattern, with the main implementation of the method using the JSON pattern with indication of nested elements. The structure is repeated for each problem-solution pair pointed out by Artificial Intelligence (AI). The inner element is a pair of start and end sub-elements, indicating where the problem is in the source code, each with the following composition: Petition 870250007435, dated 01 / 29 / 2025, p. 17 / 51 13 / 15 • Line of code, as informed in the processed code; and • On this line, the number of characters to the right from the beginning.

[043] From the point of view of reviewing and correcting source code, it is quite unproductive for problems to be pointed out far from where they occur, such as in a report separate from the source code or even in a list format at the beginning or end of the file. For this reason, the method has the inventive feature of recording each problem detection using the code review tool API (F) with identification of the exact related code segment.

[044] An important part of the invention is the generation and display of parameters that assist the developer in identifying and understanding each source code problem pointed out by Artificial Intelligence (AI). Thus, in one embodiment of the method, the prompt is written in such a way as to insert, in the meta-instruction message, information about the structuring of the responses given for each problem-solution pair in order to obtain from Artificial Intelligence (AI) the following information for each said problem-solution pair to be inserted into the comment made by the review tool (F): • Description: objective identification of the problem; • Aspect: classification of the problem using one of the following options: Code style and formatting, Code readability, Naming conventions, Comments and documentation, Code complexity, Error handling, Code efficiency, Code modularity, Security; • Explanation: the reason why the problem needs to be fixed, since often the description is not sufficient for the developer to unequivocally understand how to solve it, as typically the problematic code is not removed, but replaced with equivalent code that does not repeat the problem; • Impact: the magnitude of the consequence if it is not corrected, between the options high and low; • Frequency: counting the votes (V) of the virtual reviewers (RV), representing the spectrum between precision and comprehensiveness of the Artificial Intelligence (AI) feedback, allowing configuration by context, i.e., project by project, where the Petition 870250007435, dated 01 / 29 / 2025, page 18 / 51 14 / 15 higher frequency implies a lower chance of “false problems”, while lower frequency implies a lower chance of problems not being caught; and • Code suggestion: Code change that solves the verified problem.

[045] The parameter values ​​are obtained from Artificial Intelligence (AI) via the prompt meta-instruction, so that each virtual reviewer (VR) is guided to fill them in for each problem found.

[046] BIBLIOGRAPHIC REFERENCES • [1] US Patent 20210089992A1, “Method for automated code reviewer recommendation”. Published on 03 / 25 / 2021. Owner: NEC Laboratories America Inc. https: / / patents.google.com / patent / US20210089992A1 / • [2] US Patent 20210182031A1, “Methods and apparatus for automatic detection of software bugs”. Published on 06 / 17 / 2021. Owner: Intel Corp. https: / / patents.google.com / patent / US20210182031A1 / • [3] Vijayvergiya, Manushree, Matgorzata Salawa, Ivan Budiselic, Dan Zheng, Pascal Lamblin, Marko Ivankovic, Juanjo Carin, et al. 2024. “AI-Assisted Assessment of Coding Practices in Modern Code Review”. Proceedings of the 1st ACM International Conference on AI-Powered Software. ACM. https: / / doi.org / 10.1145 / 3664646.3665664 • [4] Froemmgen, Alexander, Jacob Austin, Peter Choy, Nimesh Ghelani, Lera Kharatyan, Gabriela Surita, Elena Khrapko, et al. 2024. “Resolving Code Review Comments with Machine Learning”. Proceedings of the 46th International Conference on Software Engineering: Software Engineering in Practice. ACM. https: / / doi.org / 10.1145 / 3639477.3639746 • [5] Amazon CodeGuru Reviewer: https: / / docs.aws.amazon.com / codeguru / latest / reviewer-ug / welcome.html • [6] Swimm: https: / / swimm.io / • [7] AI Code Review Action: https: / / github.com / marketplace / actions / ai-code-review- action • [8] Codacy: https: / / www.codacy.com / Petição 870250007435, de 29 / 01 / 2025, pág. 19 / 51 15 / 15 • [9] DeepCode AI: https: / / snyk.io / platform / deepcode-ai / •

[10] Code Climate: https: / / codeclimate.com / •

[11] Code Rabbit: https: / / www.coderabbit.ai /

Claims

1) CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS, showing a method using a manual source code review tool (F) possessing a source code repository (R), a generative Artificial Intelligence (AI), and using a computer (C) for processing, characterized by this method being carried out through four sequential steps: • Step 1 (1): Obtaining the source code files from the repository (R) by the computer (C), and processing the files;• Step 2 (2): The computer (C) constructs a prompt, written in natural language, which is then sent to the Artificial Intelligence (AI) along with the source code to be analyzed, and the prompt contains the following information: o Instruction message, describing the task to be performed, the total number of virtual reviewers (VR), indication of the recurring steps and cycles of individual review and collaborative voting (V) on the validity of each problem and the attributed impact, the programming language corresponding to the source code, and the processed source code with line information obtained in Step 1 (1); o Meta-instruction message, which describes the information exchange protocol with the Artificial Intelligence (AI); and o Randomness configuration parameters and response size parameter, adjustable at program runtime;• Step 3 (3): The prompt created in Step 2 (2) is used for Artificial Intelligence (AI) operation, the Artificial Intelligence (AI) detecting problems in the source code and generating suggested solutions for each problem verified, this is done using multiple code review agents (VR) and collaborative voting (V) to analyze each problem detected in the source code, and returning the response to the computer (C); Petition 870250007435, dated 01 / 29 / 2025, page 21 / 51 2 / 4 • Step 4 (4): The computer (C) processes the return from the Artificial Intelligence (AI) and records comments using the review tool API (F) indicating the problems pointed out line by line based on the processed file sent by the Artificial Intelligence (AI) in Step 3 (3) along with the suggested solution. 2) CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS, according to claim 1, characterized in that Step 1 (1) is performed through the following substeps: • Substep 1 (1S1): Obtaining by the computer (C) the version prior to the current one of the source code and file in unidiff format of the source code change of selected files from the repository (R); • Substep 2 (1S2): Applying by the computer (C) the changes to the source code in unidiff format of the files, obtaining a preview of the file after the change; • Substep 3 (1S3): Processing with sequential addition of source code lines to the files according to the unidiff. 3) CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS, according to claim 1, characterized in that Step 3 (3) is carried out through the following substeps: • Substep 1 (3S1): The virtual reviewers (VR), individually, examine the source code in the files, identify problems, and propose a solution for each problem pointed out in the source code; • Substep 2 (3S2): The reviewers (VR) examine each problem-solution pair pointed out by the first reviewer (VR) and perform a collaborative vote (V), expressing agreement or disagreement individually on each problem-solution pair, this being repeated for the second reviewer (VR) and so on until the last reviewer (VR); Petition 870250007435, dated 01 / 29 / 2025, page.22 / 51 3 / 4 • Substep 3 (3S3): Artificial Intelligence (AI) returns the list of problem and solution pairs from each reviewer (RV) to the computer (C); and • Substep 4 (3S4): After the return from Artificial Intelligence (AI), the computer (C) discards the problem and solution pairs that have had at least one disagreement, this substep thus eliminating the problem and solution pairs that have a high probability of being hallucinations. 4) CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS, according to claim 1, characterized by the structured output pattern of Step 3 (3) showing the problems and solutions using the JSON pattern with indication of nested elements, with structure repeating for each problem and solution pair pointed out by Artificial Intelligence (AI), and the interior element being a pair of start and end sub-elements, indicating where the problem is in the source code, each with the following composition: • Line of code, as informed in the processed code; and • In this line, number of characters to the right from the beginning. 5) CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS, according to claim 1, characterized by the prompt being written in such a way as to insert, in the meta-instruction message, information about the structuring of the responses given for each problem and solution pair to obtain from Artificial Intelligence (AI) the following information for each said problem and solution pair to be inserted in the comment made by the review tool (F): • Description: objective identification of the problem; • Aspect: classification of the problem using one of the options Code style and formatting, Code readability, Naming conventions, Comments and documentation, Code complexity, Error handling, Code efficiency, Code modularity, Security; • Explanation: reason why the problem should be corrected; Petition 870250007435, dated 01 / 29 / 2025, page.23 / 51 4 / 4 • Impact: size of the consequence if not corrected, between the options high and low; • Frequency: count of votes (V) from virtual reviewers (RV); and • Code suggestion: Code change that solves the identified problem. 6) CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS, according to claim 1, characterized in that the manual source code review tool (F) is the Azure DevOps cloud development environment and the repository (R) of the tool (F) used is Azure Repos, and the generative Artificial Intelligence (AI) is an instance of GPT-4o. 7) CODE REVIEW METHOD USING INCREMENTAL CODE PROCESSING AND ARTIFICIAL INTELLIGENCE ASSISTED WITH MULTIPLE VIRTUAL REVIEWERS, according to claim 6, characterized by the following requests being made to the Azure Repos (R) repository API: • Obtaining comments made in the Pull Request; • Recording the status of comments; • Obtaining the most recent iteration of the Pull Request, iterations being the succession of versions of the same Pull Request; • Obtaining the files changed in a specific iteration of the Pull Request; • Obtaining general information about the Pull Request; • Adding comments per line of source code, per file, and per Pull Request; • Obtaining the change in each file in unidiff format; • Obtaining the entirety of a given source code file.

Images