Ai-based active workflow agent provision system

Abstract

The present invention relates to a workflow automation technology using a large language model (LLM), and the present invention comprises: an interface module including a natural language interface for receiving a natural language query from a user, and a GUI editor which provides, to the user, a generated workflow as a UI including nodes, edges, branches and execution items and which receives, from the user, details of change in the workflow; a workflow management engine for generating a workflow including a plurality of nodes through the natural language query input from the user; and a work execution module for executing execution contents of each node according to the generated workflow. According to the present invention, an LLM agent (LLM module) is arranged as an active node (management node) on a workflow, and thus can automate a complex and dynamic work process and dynamically improving the work process.

Description

AI-based active workflow agent provision system

[0001] The present invention relates to workflow automation technology utilizing a Large Language Model (LLM), and more specifically, to an AI-based active workflow agent system that performs natural language commands, GUI-based workflow editing, JSON (JavaScript Object Notation)-based workflow definition and execution, domain-specific knowledge base (KB) integration, automatic reflection of policies and security, and automatic generation, modification, and re-execution of code.

[0002]

[0003] Workflow is a term originally used in the fields of business, economics, and administration to refer to work procedures for performing tasks. More recently, however, across all industries, including information and communications, it refers to a structured processing flow in which multiple work units are executed according to a predefined sequence, conditions, branching, or parallel structure to achieve one or more business objectives.

[0004] Such a workflow may include elements such as work nodes, connections between nodes (edges), conditional branching, loops, and parallel execution, and each work node may be configured to perform functions such as data processing, external system calls, code execution, and decision execution.

[0005] Optimization of such workflows is recognized as an important factor for improving the efficiency of business processes. To enhance work efficiency through such optimization, the workflow described in Korean Published Patent No. 10-2023-0033167 is optimized and automatically generated to improve work efficiency.

[0006] However, such existing workflow automation systems are based on static rules and operate according to sequences, conditions, and nodes clearly defined by the user. These systems are difficult to access without programming knowledge and struggle to respond flexibly to environmental changes or fluctuations in requirements. Furthermore, adding new rules or logic necessitates manual intervention, and there is a lack of immediate and automated responses in situations requiring complex external API integration, security or policy changes, or code improvements.

[0007]

[0008] The present invention has been devised to solve the aforementioned conventional problems, and aims to provide a workflow agent provision system capable of automating and dynamically improving complex and dynamic business processes by deploying an LLM agent as an active node in the workflow.

[0009] Furthermore, the present invention aims to provide a workflow agent provision system that enables non-expert users to perform workflow optimization and management without specialized knowledge through natural language queries (commands), GUI (Graphical User Interface) editing, and JSON (JavaScript Object Notation) editing.

[0010] In addition, the present invention aims to provide a workflow agent providing system capable of providing a workflow based on a professional decision-making process by utilizing domain-optimized personas and knowledge bases (KB).

[0011] Furthermore, the present invention aims to provide a system for providing a self-healing workflow agent that automatically modifies the code and performs re-execution when an error occurs during task execution by the workflow.

[0012] In addition, the present invention aims to provide a workflow agent providing system that automatically reflects security and policy changes and offers a workflow optimized for updated policies without user management.

[0013]

[0014] According to the features of the present invention for achieving the above-mentioned purpose, the present invention comprises: an interface module comprising a natural language interface that receives a natural language query from a user, and a GUI editor that provides a generated workflow to the user as a UI including nodes, edges, branches, and execution items, and receives changes to the workflow from the user; a workflow management engine that generates a workflow including a plurality of nodes through a natural language query entered by the user; and a workflow agent comprising a task execution module that executes the execution contents of each node according to the generated workflow; wherein at least one of the nodes constituting the workflow is configured as a management node in which the LLM module determines whether to modify the workflow based on the execution results of the node that has already been performed, knowledge base (KB) data, and policy information.

[0015] At this time, the above workflow may also be generated as a JSON-based data structure.

[0016] In addition, the workflow management engine may generate a workflow reflecting the change query and provide it to the user when a change query for the workflow is entered by the user through the GUI editor.

[0017] Additionally, the LLM module may modify the workflow by reflecting one or more of path resetting, code modification, and authentication node insertion when a security event occurs during the execution of the workflow, including a change in API response format, performance degradation, policy change, and security requirement.

[0018] In addition, the above workflow management engine may set a persona for the above LLM module based on domain-specific knowledge base (KB) data.

[0019] In addition, the detailed role of the LLM module executing the management node may be set for the above-mentioned Persona according to the execution content of the nodes before and after the management node in the workflow.

[0020] In addition, if an error occurs during node execution, the above LLM module may analyze JSON-based error logs and, based on the analysis results, modify code snippets or generate alternative algorithms to instruct re-execution.

[0021] In addition, the workflow is configured to include action nodes including an external API integration node, a data processing node, a file input / output node, a DB access node, and a code execution node; and the LLM module may analyze the execution results and status of the action nodes to perform parallel processing optimization, conditional branch path changes, and adjustment of the number of iterations.

[0022] In addition, the above workflow agent may be configured to include a management module that supports authentication / authorization procedures through Role-Based Access Control (RBAC) and Single Sign-On (SSO).

[0023] In addition, the LLM module may execute security / policy processes including sensitive information masking, authentication node insertion, and API key management by referring to policy information.

[0024] In addition, the above workflow agent may be configured to include a monitoring module that records executed commands, workflow change history, execution results, and LLM decision history as JSON logs, and performs subsequent audit trails and root cause analysis through said records.

[0025] At this time, the LLM module provides the user with improvement proposals, including pipeline structure improvement, algorithm replacement, and parallel processing strategy optimization, based on long-term accumulated execution records and performance indicators through the GUI editor; and the workflow management engine may generate and apply a workflow reflecting the improvement proposal when the user selects the improvement proposal through the GUI.

[0026] In addition, the above GUI editor may provide standard nodes, which are domain-specific common business logic, in the form of templates, thereby enabling the rapid creation of workflows by combining or modifying the standard nodes.

[0027] In addition, the above LLM module may refer to knowledge base (KB) data including policy information, regulatory information, industry standards, and quality guidelines to dynamically identify policy and regulatory changes occurring during workflow execution, and modify the workflow structure or code logic to conform to such changes.

[0028] In addition, the workflow agent supports a container-based deployment environment and can variably adjust resource allocation in response to the increase in parallel processing nodes and the demand for improved computational performance.

[0029] And the interface module further comprises a query pattern learning unit that stores and learns natural language queries input by a user; the query pattern learning unit manages the input queries by workflow as query groups, and compares the input time of a newly input query group with the input time of previously learned query groups; if the input time of the newly input query group by the user is the minimum time, the workflow generated by the newly input query group by the user may be used as training data to perform reinforcement learning or self-supervised learning of the user's input query pattern.

[0030] In addition, the query pattern learning unit may be configured to include a query synonym learning unit that learns query groups input by a user and generates a query standardization table.

[0031] And the above query standardization table may be a table in which one or more input queries are corrected and matched to a standard query to be executed.

[0032] In addition, the above query synonym learning unit may generate a query standardization table by comparing and analyzing query groups having the same workflow, and, regarding the execution result of an input query, if a re-entered query exists, learning the previously entered query as the query to be corrected and the re-entered query as the standard query.

[0033]

[0034] The workflow agent providing system according to the present invention, as described above, can be expected to have the following effects.

[0035] In other words, in the present invention, an LLM agent (LLM module) is deployed as an active node (management node) on the workflow, thereby enabling the automation of complex and dynamic business processes and the dynamic improvement of such processes.

[0036] Furthermore, the present invention has the effect of enabling even non-expert users to optimize and manage workflows through general conversational input without specialized knowledge, via natural language queries (commands), GUI (Graphical User Interface) editing, and JSON (JavaScript Object Notation) editing.

[0037] In addition, the present invention utilizes personas and knowledge bases (KB) optimized for each domain, workflow, and individual node to provide a workflow based on a professional decision-making process, thereby improving the performance of workflow optimization.

[0038] In addition, in the present invention, if an error occurs during task execution by the workflow, the code is automatically modified and re-executed to enable self-healing, thereby improving user convenience.

[0039] In addition, the present invention has the effect of automatically reflecting security and policy changes, thereby providing a workflow optimized for updated policies without user management.

[0040] Furthermore, in the present invention, the workflow agent learns the causal relationship between consecutive queries input by a user and the execution results (workflow) that are executed, thereby reducing the gap between the user's input query and the task (workflow) intended by the user, and thus has the effect of enabling the generation of a workflow tailored to the user's intent.

[0041]

[0042] FIG. 1 is a block diagram illustrating the configuration of a workflow agent providing system according to a specific embodiment of the present invention.

[0043] FIG. 2 is an exemplary diagram illustrating a node configuration example of a workflow generated by a workflow agent according to the present invention.

[0044] FIG. 3 is a flowchart illustrating an example of the execution process of a management node of an LLM module constituting the present invention.

[0045] FIG. 4 is an illustrative diagram showing an example of a persona set by a workflow agent according to the present invention.

[0046] FIG. 5 is an example diagram conceptualizing and illustrating a query entered by a user and a workflow generated by it in the present invention.

[0047]

[0048] The present invention relates to a workflow automation technology utilizing a Large Language Model (LLM), comprising: an interface module comprising a natural language interface that receives a natural language query from a user, and a GUI editor that provides a generated workflow to the user as a UI including nodes, edges, branches, and execution items, and receives changes to the workflow from the user; a workflow management engine that generates a workflow including multiple nodes through a natural language query entered by the user; and a workflow agent comprising a task execution module that executes the execution contents of each node according to the generated workflow; wherein at least one of the nodes constituting the workflow is composed of a management node that determines whether to modify the workflow based on the execution results of a node where the LLM module has already performed, knowledge base (KB) data, and policy information; and the workflow is generated as a JSON-based data structure.

[0049]

[0050] Hereinafter, we will examine an AI-based active workflow agent providing system according to a specific embodiment of the present invention with reference to the attached drawings.

[0051] Before proceeding with the explanation, the effects, features, and methods for achieving the present invention will become clear from the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but can be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims.

[0052] In describing the embodiments of the present invention, if it is determined that a detailed description of known functions or configurations may unnecessarily obscure the essence of the invention, such detailed description will be omitted. Furthermore, the terms described below are defined considering the functions in the embodiments of the present invention, and these may vary depending on the intentions or conventions of the user or operator. Therefore, such definitions should be based on the content throughout this specification.

[0053] Combinations of each block of the attached block diagram and each step of the flowchart may be executed by computer program instructions (execution engine), and since these computer program instructions may be loaded into the processor of a general-purpose computer, a special-purpose computer, or other programmable data processing equipment, the instructions executed through the processor of the computer or other programmable data processing equipment create a means to perform the functions described in each block of the block diagram or each step of the flowchart.

[0054] Since these computer program instructions may be stored in computer-available or computer-readable memory that can be directed toward a computer or other programmable data processing equipment to implement a function in a specific way, the instructions stored in said computer-available or computer-readable memory may also be used to produce a manufactured item containing instruction means that perform the function described in each block of a block diagram or each step of a flowchart.

[0055] And since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that execute a computer or other programmable data processing equipment by performing a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in each block of the block diagram and each step of the flowchart.

[0056] Additionally, each block or each step may represent a module, segment, or part of code containing one or more executable instructions for executing specific logical functions, and in some alternative embodiments, the functions mentioned in the blocks or steps may occur out of order.

[0057] In other words, the two blocks or steps described can actually be performed substantially simultaneously, and can also be performed in the reverse order of their corresponding functions as needed.

[0058]

[0059] FIG. 1 is a block diagram illustrating the configuration of a workflow agent providing system according to a specific embodiment of the present invention, FIG. 2 is an example diagram illustrating an example of a node configuration of a workflow generated by a workflow agent according to the present invention, FIG. 3 is a flowchart illustrating an example of the execution process of a management node of an LLM module constituting the present invention, FIG. 4 is an example diagram illustrating an example of a persona set by a workflow agent according to the present invention, and FIG. 5 is an example diagram conceptualizing a query entered by a user and a workflow generated by it in the present invention.

[0060] First, as illustrated in FIG. 1, the workflow agent providing system according to the present invention comprises a user terminal (100) and a workflow agent (200) that processes natural language queries input from the user terminal and workflow creation and editing, etc.

[0061] Here, the workflow agent (200) is configured to include an interface module (210), a workflow management engine (220), an LLM module (240), an execution module (230), a management module (250), and a monitoring module (260).

[0062] The above interface module (210) is a part that receives natural language queries from a user, provides a generated workflow to the user, and receives various input commands from the user, and is configured to include a natural language interface (211), a GUI editor (213), and a query pattern learning unit (215).

[0063] Here, the natural language interface (211) is a part that receives a natural language query from a user and interprets the natural language input using an LLM model, and the natural language query can be input through various means such as text, voice, etc.

[0064] At this time, the LLM model for interpreting the natural language query may use the LLM module (140) described later, and the natural language interface (211) may use a separate LLM model.

[0065] And the above GUI editor (213) provides the user with the generated workflow in an intuitive UI form and receives the user's requirements, such as modifications or additions, through the provided UI. It is desirable that the provided UI includes nodes, edges, branches, and execution items that constitute the workflow.

[0066] That is, the generated workflow is visually provided to the user through a GUI editor (213), and the user can input the addition, deletion, change of connection relationships, and modification of execution conditions of nodes through GUI operations.

[0067] In this case, the workflow management engine (220) reflects the changes entered through the GUI editor (213), updates the modified workflow into a JSON data structure, and provides it back to the user.

[0068] Accordingly, in the present invention, natural language input, GUI editing, and JSON data structures operate in conjunction, so that even non-expert users can create and manage workflows without programming knowledge.

[0069] Meanwhile, the above query pattern learning unit (215) identifies the natural language expression characteristics of the user and, considering the user's linguistic characteristics, enables the workflow management engine (220) to generate an optimal workflow. The specific details will be explained again when describing FIG. 5.

[0070] Next, the workflow management engine (220) is a part that creates and manages a workflow including multiple nodes through a natural language query entered by the user.

[0071] At this time, the workflow generated by the workflow management engine (220) is generated as a JSON-based data structure and consists of various nodes as shown in FIG. 2.

[0072] The nodes constituting the above workflow can be classified according to various criteria and can be broadly configured to include Trigger nodes, Action nodes, Function nodes, Integration nodes, etc. In the case of Action nodes, they can be configured to include external API integration nodes, data processing nodes, file I / O nodes, DB access nodes, and code execution nodes.

[0073] In particular, the workflow according to the present invention is configured to include a management node, as shown in FIG. 2.

[0074] The above management node is a node that executes optimization by managing the workflow in real time through the above LLM module (240), and will be explained in detail later when describing the function of the LLM module (240).

[0075] Meanwhile, the workflow management engine (220) can perform modification or regeneration of the generated workflow. For example, if a user inputs a change query containing modification or change details regarding the presented workflow through the GUI editor (213), the workflow with the change query reflected is generated and provided to or executed by the user.

[0076] Next, the above-mentioned task execution module (230) is a part that executes the execution content of each node according to the generated workflow, and can perform the execution content by connecting to an external server (execution server, 310, 320) or database according to the execution content of the node.

[0077] Meanwhile, the above LLM module (240) is implemented as a large-scale language model and is a part that performs the role of a major AI agent according to the present invention, and basically runs a management node.

[0078] At this time, the management node is a node that determines the optimality of the workflow currently being executed based on the execution results of previously performed nodes, knowledge base (KB) data, and policy information, and evaluates the necessity of supplementing the workflow according to the determination result.

[0079] As described above, the LLM module (240) executes a management node to manage the optimization of the workflow, and FIG. 3 illustrates the process of the LLM module supplementing the workflow in the event of a supplementary event caused by internal or external factors and in the event of an execution error.

[0080] As illustrated in FIG. 3, the LLM module (240) manages the node execution content during the workflow execution process (S100) and checks whether a supplementary event occurs (S200).

[0081] At this time, a supplementary event is an event that requires supplementation of the workflow and may be caused by various factors, for example, when code modification is required because the API response format is changed during the execution of the workflow; when path reconfiguration is required because execution performance is degraded; or when it is necessary to insert an additional authentication node due to external policy changes or security requirements.

[0082] In this way, when a supplement event occurs, the LLM module (240) modifies the workflow and executes it by modifying the code (S220), resetting the path (S210), adding an authentication node (S230), etc., according to the content of the supplement request.

[0083] Meanwhile, the LLM module (240) modifies the workflow for the error even if an error occurs in the node execution process during the workflow execution process, thereby maintaining optimization without user intervention.

[0084] To this end, the LLM module (240) analyzes a JSON-based error log (S310) when an error occurs during node execution (S300).

[0085] And, if the LLM module (240) determines that a specific part causes an execution error based on the analysis results, it modifies the code snippet (S320), or if it determines that there is a problem with the algorithm configuration, it generates an alternative algorithm (S330) to re-execute.

[0086] Accordingly, the LLM module (240) can analyze JSON-based error logs and perform a self-healing function in response to the cause of the error.

[0087] Furthermore, the LLM module (240) can analyze the execution results and status of the action node to perform parallel processing optimization, change the conditional branch path, and adjust the number of iterations.

[0088] And the above LLM module (240) may also execute security / policy tasks including sensitive information masking, authentication node insertion, and API key management by referring to policy information provided by the domain.

[0089] Meanwhile, in order to improve the execution effect when the above LLM module (240) executes the management node, a persona suitable for the execution task may be set as shown in FIG. 4.

[0090] At this time, the above persona may be configured based on domain-specific knowledge base (KB) data or based on a generated workflow.

[0091] In other words, by analyzing the knowledge base (KB) data provided in the domain where the workflow is applied, the relevant business field can be calculated, and accordingly, experts in that field (MES experts, DevOps experts, HR experts, education experts, and R&D experts) can be established.

[0092] However, in the present invention, the LLM module (240) manages the workflow in real time through a management node during workflow execution, and since the detailed role of the persona may vary depending on the execution node, the detailed role of the persona may be set according to the execution content of the nodes before and after the management node on the workflow.

[0093] For example, assuming that the workflow being executed pertains to corporate human resource management, specifically the task of establishing improved future hiring criteria by comparing and analyzing employees' past hiring standards and scores with their current performance evaluation results, while 'HR expert' is a persona that fits the overall workflow, the persona 'HR department performance evaluation expert' is more suitable when executing nodes related to analyzing performance evaluations, and the persona 'HR department recruitment expert' is more suitable when executing nodes related to establishing hiring criteria. Therefore, the present invention can derive more optimized results by setting such detailed personas for each node.

[0094] Meanwhile, although it was previously stated that the Persona setting for the LLM module (240) can be assigned by the workflow management engine (220), it is also possible for the LLM module (240) to assign it itself, or for it to be assigned by another component constituting the workflow agent (200) according to the present invention.

[0095] Furthermore, although the workflow management engine (220) and the LLM module (240) have been described as separate components in the present invention, since the LLM module (240) can also generate a workflow, it is possible for the LLM module (240) to perform the functions of the workflow management engine (220).

[0096] The above management module (250) is a part that supports authentication and authorization procedures through role-based access control (RBAC) and single sign-on (SSO).

[0097] And the above monitoring module (260) records the executed commands, workflow change history, execution results, and LLM decision history as a log in JSON format, and performs subsequent audit trail and problem cause analysis through said record.

[0098] Meanwhile, as described above, the query pattern learning unit (215) learns the user's input query pattern through the collected user input queries. At this time, the query pattern refers to a series of queries entered by the user to execute a specific task.

[0099] In this case, it is desirable for the above query pattern to be learned by grouping the input queries into workflow (task) units.

[0100] To this end, the query pattern learning unit (215) groups and classifies the queries entered by the user according to the time between query inputs and the similarity of the workflow generated according to the query inputs, and the query pattern learning unit (215) can learn query patterns for each query group.

[0101] And the above query pattern learning unit (215) learns the input cycle of query groups classified by work unit through the collected user input queries.

[0102] In this way, when a specific query is entered by a user, the assist algorithm learned by the query pattern learning unit (215) can produce a prediction result regarding the queries (query group) to be followed by the query, the workflow to be generated through this, and the period for the re-entry of the query group.

[0103] Furthermore, the above query pattern learning unit (215) can enable the queries constituting a specific query group to be learned in an optimized state.

[0104] For example, even though the same user performs the same task (workflow), query groups can be configured with different queries depending on the situation.

[0105] At this time, the query pattern learning unit (215) compares the input time of a query group newly entered by the user with the input time of previously learned query groups, and if the input time of the query group newly entered by the user is the minimum time, the user's input query pattern can be reinforced learning or self-supervised learning by using the workflow generated by the query group newly entered by the user as learning data.

[0106] Here, the input time of a query group refers to the time required for natural language queries of a query group to be input until the user creates a completed (satisfactory) workflow.

[0107] That is, when creating a workflow, the fact that the input time of the input queries is short means that the same workflow is created with efficient query input, and the query pattern learning unit (215) can learn a more efficient query group for repetitive tasks (workflow) by reinforcing learning the workflow and query group in such cases.

[0108] For example, as shown in FIG. 5, if we assume that to generate the same final result 'WF(Work Flow)4', query group 1 inputs four queries A1, A2, A3, and A4 to generate the final result 'WF4' through the results of WF1, WF2, WF3, and WF4, respectively, and query group 2 inputs three queries A1, A3, and A4 to generate the final result 'WF4' through the results of WF1, WF3, and WF4, respectively, then the execution schedule of query group 2 can be determined to be more efficient.

[0109] Accordingly, the user is more likely to configure the query group as in the execution schedule of query group 2, and accordingly, the query pattern learning unit (215) according to the present invention can derive a more efficient query pattern as a prediction result by reinforcing learning (or self-supervised learning) the query pattern and workflow of query group 2.

[0110] Next, the query pattern learning unit (215) according to the present invention may be configured to self-correct errors in queries that are frequently entered incorrectly by the user and derive an execution result.

[0111] The workflow agent (200) according to the present invention receives natural language queries by applying a large language model. Since the user's concept of a specific word (or sentence) and the concept of the word (or sentence) recognized by the artificial intelligence algorithm are different, results different from the user's intention may occur, and there are many cases where the same word or sentence is repeated for each individual.

[0112] This is due to the fact that while LLM concepts are learned through the average language perception of large groups of users, individual users establish concepts of language based on their personal experiences or habits, which are not easily changed.

[0113] The query pattern learning unit (215) according to the present invention can utilize a query group and a result generated therefrom in order to correct the query that the user repeatedly inputs incorrectly and execute the result.

[0114] Specifically, the query pattern learning unit (215) may be configured to include a query synonym learning unit (not shown) that learns query groups input by a user and generates a query standardization table.

[0115] In this case, the above query standardization table refers to a table in which one or more input queries are corrected and matched to the standard query to be modified.

[0116] In other words, the aforementioned query standardization table refers to the matching of incorrectly entered user queries with standard queries that must be input into the large language model to derive execution results that match the user's intent.

[0117] In order to generate the above query standardization table, the query synonym learning unit compares and analyzes query groups having identical execution results (workflows), and regarding the execution result of an input query, if a re-entered query exists, learns the previously entered query as the query to be corrected and the re-entered query as the standard query to generate the query standardization table.

[0118] For example, as shown in FIG. 5, if we assume that in order to execute the same final result 'WF4', query group 2 inputs queries A1, A3, and A4 to execute the final result 'WF4' through the results of WF1, WF3, and WF4, respectively, and query group 3 inputs queries A0, A1, A3, and A4 to execute the final result 'WF4' through the results of WF0, WF1, WF3, and WF4, respectively, then query A0 of query group 3 can be determined to be a query incorrectly entered by the user who intended the execution result WF1.

[0119] In this case, the query standardization table can be organized by matching query A0 with A1 as the standard query, and the query pattern learning unit (215) can correct and execute it as the standard query A1 when query A0 is input from the user.

[0120] Accordingly, the present invention makes it possible to change queries that are frequently entered incorrectly due to the user's language habits into corrected queries that match the user's intent and execute them.

[0121]

[0122] Below, we will examine application examples in which the workflow agent providing system according to the present invention is applied to various business fields.

[0123] First, when a workflow for global manufacturing process (MES) optimization and predictive maintenance is created through the workflow agent providing system according to the present invention (Application Example 1), the execution process is as follows when the business requirements are: ① real-time collection of sensor data from three factories worldwide, quality inspection, and production performance analysis; ② automatic readjustment of production parameters when the defect rate rises by more than 2%; ③ automatic execution of an emergency contract with an alternative supplier when there is a shortage after monitoring raw material inventory; ④ automatic generation of machine learning code by an LLM node based on production / quality data from the past month at regular intervals (end of the month), demand forecasting for the next quarter, and production schedule reallocation; ⑤ change to specific advanced nodes is only possible by an administrator through internal MES standards, GMP (Good Manufacturing Practice) documents, reference to international quality regulations KB, and authorization management (RBAC); and ⑥ automatic reflection of changes to authentication / security policies (e.g., insertion of a new API authentication token processing node).

[0124]

[0125] (1) Natural language commands

[0126] We collect sensor data from factories worldwide every hour for quality inspections, modify equipment parameters if the defect rate exceeds 2%, contract with alternative suppliers in case of raw material shortages, and update production planning forecast models based on historical data at the end of every month.

[0127]

[0128] (2) LLM Analysis & Workflow Creation

[0129] A workflow is created for the LLM (MES Expert Persona) node that includes sensor data collection nodes (each factory API), quality inspection code nodes, parameter readjustment nodes, inventory verification and order nodes, machine learning prediction nodes, etc.

[0130] In this case, the above workflow actively utilizes node templates for MES and the GMP / Quality Regulation Knowledge Base (KB). Additionally, if there are policy changes such as 'strengthening certification procedures,' an LLM module is inserted as an authentication node.

[0131]

[0132] (3) GUI editing

[0133] The Global Production Management General Manager defines node placement and parallel processing groups in the GUI, and connects conditional branches to inventory lookup nodes and supplier contract nodes.

[0134]

[0135] (4) Dynamic decision-making during execution

[0136] When the sensor format is changed, the LLM module analyzes the error log, modifies the parsing code, and re-executes it. In this case, if the failure rate spikes, an alternative algorithm (another parameter tuning method) is proposed.

[0137]

[0138] (5) Performance monitoring / logging

[0139] It audits JSON log records for all executions and controls scaling out to Kubernetes based on whether it runs stably even with large-scale parallel processing (3 factories, hundreds of sensors).

[0140]

[0141] (6) Long-term analysis

[0142] After accumulating one month of data, the LLM module automatically generates code through machine learning, and the prediction model is improved. Based on the prediction results, it proposes adding nodes through production schedule reallocation, and these changes are implemented after approval via the GUI editor.

[0143]

[0144] Next, when a workflow for improving the advanced DevOps pipeline and continuous code quality management is created through the workflow agent providing system according to the present invention (Application Example 2), the business requirements are as follows: ① Automatic build and unit / integration test execution when a commit occurs for dozens of microservice code repositories ② Scanning container images for security vulnerabilities and if quality is substandard, the LLM proposes code improvement plans ③ Test coverage and performance measurement → if below the threshold, routing to a Slack notification node instead of Canary deployment ④ Referencing internal DevOps best practice documents and cloud deployment policy KB ⑤ Only DevOps administrators can modify specific advanced nodes (deployment-related nodes) via RBAC. The execution process is as follows.

[0145]

[0146] (1) Natural language commands

[0147] When a commit occurs in each service, a build, test, and vulnerability scan are performed; if quality criteria are met, a Canary deployment is initiated, otherwise a Slack notification is sent. Every two weeks, an LLM analyzes code quality patterns and suggests improvements to the pipeline structure.

[0148]

[0149] (2) LLM Analysis & Workflow Creation

[0150] The LLM module (DevOps engineer persona) creates a workflow including a build node, test node, vulnerability scan node, performance analysis node, Canary deployment node, and Slack alert node. At this time, a KB for the DevOps guide is referenced, and an authentication token management node may be inserted.

[0151]

[0152] (3) GUI editing

[0153] DevOps administrators can configure parallel build nodes (simultaneous processing of multiple microservices) in the GUI editor, and when a Canary deployment node is clicked, a path can be set to an “engineer” persona and a specific API specification document.

[0154]

[0155] (4) Dynamic decision-making

[0156] If a test fails, the LLM node analyzes the logs, generates code modifications (e.g., code to improve test coverage), and re-runs the test.

[0157] In addition, if performance degradation occurs, an alternative setting (increased resource allocation) is automatically suggested.

[0158]

[0159] (5) Long-term improvement

[0160] After two weeks, the LLM node detects repetitive error patterns and proposes pipeline structure optimization (removal of unnecessary test steps, optimization of parallel processing), which is then applied after approval via the GUI editor. Additionally, the JSON log records are audit-traceed.

[0161]

[0162] Meanwhile, when a workflow for large-scale research (R&D) data analysis and intelligent pipeline refactoring is created through the workflow agent providing system according to the present invention (Application Example 3), the execution process is as follows when the business requirements are: ① daily collection of experimental data of hundreds of GB in size from an HPC (High Performance Computing) cluster, ② Spark-based preprocessing, machine learning outlier detection, automatic generation of result graphs and PDF reports, ③ monthly analysis of results by an LLM node, automatic pipeline refactoring by generating new efficient algorithm code, and ④ automatic application of research ethics, data privacy regulations (KB), and masking of sensitive information.

[0163] (1) Natural language commands

[0164] Daily HPC cluster data preprocessing, outlier detection, and generation of graphs and PDF reports. Monthly performance analysis leads to LLM automatically improving the pipeline with more efficient algorithm code.

[0165]

[0166] (2) LLM Analysis & Workflow Creation

[0167] The LLM module (data scientist persona) creates a workflow including a data collection node, a Spark preprocessing code node, a machine learning model node, a graph generation node, a PDF report node, and a performance monitor node. At this time, the Research Ethics KB is referenced, and a sensitive data masking node may be inserted.

[0168]

[0169] (3) GUI editing

[0170] Research institute data scientists can add iterative execution (daily) and monthly performance analysis nodes (conditional branching) in the GUI editor. Additionally, options such as “Prioritize high-performance algorithms” can be selected in the LLM module’s prompt panel.

[0171]

[0172] (4) Dynamic decision-making

[0173] When the data format changes on a specific day, the LLM module detects the error and may re-execute after improving the code. Additionally, if a machine learning model performance degradation is detected, the LLM module may generate and propose code for a different algorithm (e.g., Random Forest → XGBoost).

[0174]

[0175] (5) Long-term improvement

[0176] After one month, the LLM module iteratively analyzes patterns, and suggestions such as pipeline node rearrangement (parallelizing preprocessing steps) and more efficient API call structures can be proposed. After approval via the GUI editor, JSON updates and log audit trails are performed.

[0177]

[0178] Next, when a workflow for HR global workforce management and automated training curriculum is created through the workflow agent providing system according to the present invention (Application Example 4), the execution process is as follows when the business requirements are: ① collection of a list of new hires from each country branch at the end of each month, resume skill analysis, and recommendation of customized training based on missing skills; ② compliance with regional HR regulations and personal information protection policies, and masking of sensitive data; and ③ collection of performance data every six months, after which the LLM node regenerates the training curriculum code and automatically reflects changes in international HR policies (e.g., insertion of additional authentication nodes).

[0179]

[0180] (1) Natural language commands

[0181] At the end of each month, gather new hire information from each branch, analyze resumes, assign training for 부족한 skills, protect sensitive information, and redesign the curriculum after reviewing performance after six months.

[0182]

[0183] (2) LLM Analysis & Workflow Creation

[0184] The LLM module (HR expert persona) generates a workflow that includes HR API nodes, resume parsing code nodes, training recommendation nodes, personal information masking nodes, performance evaluation nodes, curriculum redesign nodes, etc. At this time, the International HR Policy Information Base (KB) is referenced, and authentication nodes, etc., can be inserted by analyzing security policy documents.

[0185]

[0186] (3) GUI editing

[0187] The HR manager can configure conditional branching (assigning specific training courses when specific skills are lacking), etc., using the GUI editor.

[0188] When the management node is clicked, you can manually set the “Expert” persona and specify the “document KB path” in the prompt panel.

[0189]

[0190] (4) Dynamic decision-making

[0191] If the resume format for a specific region changes, the LLM module modifies the parsing code and re-executes it.

[0192] In addition, after 6 months, the performance data analysis can propose and implement structural changes to the LLM module's educational curriculum (e.g., a list of educational courses expressed in code).

[0193]

[0194] (5) Policy reflection

[0195] When updating personal information protection regulations, the LLM module changes the masking node parameters and applies them after user approval through the GUI editor.

[0196]

[0197] Finally, when a workflow for proposing a customized learning path in the field of education and automatic difficulty adjustment is created through the workflow agent providing system according to the present invention (Application Example 5), the execution process is as follows when the work requirements are: ① weekly record of students' lecture attendance and analysis of quiz scores, ② recommendation of additional learning materials tailored to deficient concepts and automatic regeneration of quizzes by difficulty level, ③ generation of curriculum reallocation code by the LLM node analyzing overall performance after one semester, and ④ reference to education policy documents (KB) and compliance with accessibility guidelines.

[0198]

[0199] (1) Natural language commands

[0200] Weekly analysis of individual student quiz scores and viewing history to recommend supplementary materials on 부족한 concepts and regenerate quizzes with adjusted difficulty. Curriculum reallocation based on performance reports after one semester.

[0201]

[0202] (2) LLM Analysis & Workflow Creation

[0203] The LLM module (education expert persona) creates a workflow including LMS API call nodes, quiz generation code nodes, material recommendation nodes, difficulty adjustment parameter nodes, and curriculum rearrangement nodes. At this time, educational policy KBs (learning ethics, accessibility) may be referenced, and authentication nodes may be inserted if necessary.

[0204]

[0205] (3) GUI editing

[0206] The educational institution e-learning manager can connect nodes for recurring execution (weekly) and conditional branching (additional materials for students below a grade threshold) in the GUI editor, and options such as “beginner-friendly difficulty” can be selected in the LLM module prompt panel.

[0207]

[0208] (4) Dynamic decision-making

[0209] When the LMS response format changes, the LLM module can improve the parsing code and re-execute it.

[0210] In addition, if difficulty with repetitive learning is detected for a specific group of students, the LLM module can suggest generating additional support material code nodes.

[0211]

[0212] (5) Long-term improvement

[0213] After one semester, the LLM module can propose curriculum rearrangement code based on accumulated data, and the workflow can be updated via the GUI editor after user approval.

[0214]

[0215] The rights of the present invention are not limited to the embodiments described above but are defined by the claims, and it is obvious that a person skilled in the art may make various modifications and adaptations within the scope of the rights described in the claims.

[0216]

[0217] The present invention relates to a workflow automation technology utilizing a Large Language Model (LLM). According to the present invention, an LLM agent (LLM module) is deployed as an active node (management node) on the workflow, thereby enabling the automation of complex and dynamic business processes and dynamic improvement.

Claims

1. An interface module comprising a natural language interface that receives a natural language query from a user, and a GUI editor that provides a generated workflow to the user as a UI including nodes, edges, branches, and execution items, and receives changes to the workflow from the user; A workflow management engine that generates a workflow including multiple nodes through a natural language query entered by a user; and A workflow agent configured to include a task execution module that executes the execution contents of each node according to the above-mentioned generated workflow; and One or more of the nodes constituting the above workflow are, An AI-based active workflow agent providing system characterized by comprising an LLM module composed of a management node that determines whether to modify a workflow based on the execution results of a previously performed node, knowledge base (KB) data, and policy information.

2. In Paragraph 1, The above workflow is, AI-based active workflow agent provisioning system characterized by being generated with a JSON-based data structure.

3. In Paragraph 2, The above workflow management engine is, An AI-based active workflow agent providing system characterized by generating and providing a workflow reflecting a change query to the user when a change query for the workflow is entered by a user through the GUI editor.

4. In Paragraph 3, The above LLM module is, An AI-based active workflow agent providing system characterized by modifying the workflow by reflecting one or more of path resetting, code modification, and authentication node insertion when a supplementary event, including a change in API response format, performance degradation, policy change, and security requirement, occurs during the execution of the workflow.

5. In Paragraph 4, The above The workflow management engine is, AI-based active workflow agent providing system characterized by setting a persona for the above LLM module according to domain-specific knowledge base (KB) data.

6. In Paragraph 5, The above LLM module is, An AI-based active workflow agent providing system characterized by analyzing JSON-based error logs when an error occurs during node execution, modifying code snippets or generating alternative algorithms based on the analysis results, and instructing re-execution.

7. In Paragraph 6, The above workflow is, It is configured to include action nodes including an external API integration node, a data processing node, a file input / output node, a DB access node, and a code execution node; The above LLM module is, An AI-based active workflow agent providing system characterized by analyzing the execution results and status of the above action nodes to perform parallel processing optimization, conditional branch path changes, and adjustment of the number of iterations.

8. In Paragraph 7, The above workflow agent is, An AI-based active workflow agent providing system characterized by being configured to include a management module that supports authentication / authorization procedures through Role-Based Access Control (RBAC) and Single Sign-On (SSO).

9. In Paragraph 8, The above LLM module is, An AI-based active workflow agent providing system characterized by executing security / policy processes, including sensitive information masking, authentication node insertion, and API key management, by referring to policy information.

10. In Paragraph 9, The above workflow agent is, An AI-based active workflow agent providing system characterized by being configured to include a monitoring module that records executed commands, workflow change history, execution results, and LLM decision history as JSON logs, and performs subsequent audit trails and root cause analysis through said records.

11. In Paragraph 10, The above LLM module is, Based on long-term accumulated execution records and performance indicators, improvement plans including pipeline structure improvement, algorithm replacement, and parallel processing strategy optimization are provided to the user through the aforementioned GUI editor; The above workflow management engine is, An AI-based active workflow agent providing system characterized by creating and applying a workflow reflecting the improvement plan when a user selects the improvement plan through a GUI.

12. In any one of paragraphs 1 to 11, The above GUI editor is, An AI-based active workflow agent providing system characterized by providing standard nodes, which are domain-specific common business logics, in the form of templates, thereby enabling the rapid creation of workflows by combining or modifying the standard nodes.

13. In Paragraph 12, The above LLM module is, An AI-based active workflow agent providing system characterized by referencing knowledge base (KB) data including policy information, regulatory information, industry standards, and quality guidelines to dynamically identify policy and regulatory changes occurring during workflow execution and modifying the workflow structure or code logic to conform to the changes.

14. In Paragraph 13, The above workflow agent is, An AI-based active workflow agent providing system that supports a container-based deployment environment and is characterized by variable resource allocation in response to the increase in parallel processing nodes and the demand for improved computational performance.

Images