Intelligent quotation collaborative decision-making platform for enterprise products

Abstract

The invention provides an enterprise product intelligent quotation collaborative decision-making platform. A platform architecture comprises a user interaction layer, a data processing layer and a decision-making layer, and platform functions are realized by relying on an intelligent agent. When a user releases a product quotation task, an authority management agent verifies the identity and access authority of the user, and a user interaction and demand analysis agent in a user interaction layer converts the natural language demand of the user into a structured instruction; the data processing layer receives an instruction, basic data and a special cost agent in the layer pull data from an external system through a cross-system calling agent, and a price calculation agent integrates the data through multi-objective optimization to generate candidate quoted prices; and finally, a man-machine cooperation and intelligent decision-making auxiliary agent in a decision-making layer displays all candidate schemes, and an expert performs fine adjustment to select an optimal quotation. According to the invention, rapid and accurate quotation is realized for enterprise products, the quotation model is continuously optimized through continuous feedback and self-learning, and finally a highly autonomous intelligent quotation system is formed.

Description

Technical Field

[0001] The present invention involves natural language processing (NLP), generative large models, knowledge graphs, reinforcement learning, big data analysis, data mining, multimodal fusion, virtual simulation, API security, and distributed transaction processing technology. It is applied to the field of enterprise product quotation and is a collaborative decision-making platform for intelligent quotation of enterprise products. Background Art

[0002] With the rapid development of the information age, the demand for timeliness, high quality and professionalism in the field of corporate product quotations is increasing.

[0003] Traditional enterprise product quotation models suffer from serious deficiencies in data management. Key data is often stored in multiple systems, lacking centralized management and standardized interfaces between them. This creates information silos, leading to slow data updates and inconsistent versions. This prevents quoters from obtaining the latest and most accurate information when formulating quotes, impacting the accuracy and effectiveness of their decision-making. Furthermore, traditional quotation models often rely on fixed rules and manual experience, lacking real-time feedback on market changes. Faced with factors like fluctuating raw material prices and volatile customer demands, they are unable to rapidly and dynamically integrate data and conduct multi-angle analysis. Consequently, the resulting quotation proposals are based solely on historical data and static analysis, becoming disconnected from actual conditions. Furthermore, the quotation decision-making process is simplistic, lacking the ability to generate multiple proposals in parallel and perform virtual simulations. This makes it impossible to identify potential optimization opportunities through comparative analysis of different proposals, resulting in final quotes that fail to meet actual needs and deviate from market conditions. Overall, traditional quotation systems lack system collaboration, information sharing, and dynamic decision-making, making it difficult for enterprises to develop efficient and accurate product quotation strategies in a complex and rapidly changing market, leading to missed market opportunities. Summary of the Invention

[0004] The objective of the present invention is to propose an enterprise product intelligent quotation collaborative decision-making platform to solve the problems raised in the above background technology.

[0005] The present invention proposes an enterprise product intelligent quotation collaborative decision-making platform. The platform architecture includes a user interaction layer, a data processing layer, and a decision-making layer. The functions of each layer in the platform are realized by the collaboration of multiple intelligent agents, among which:

[0006] User interaction and demand analysis agents are deployed in the user interaction layer; basic data agents, special expense agents, and price calculation agents are deployed in the data processing layer; human-computer collaboration and intelligent decision-making assistance agents are deployed in the decision-making layer; cross-system call agents and permission management agents are deployed in the platform and flexibly called by other agents according to business needs.

[0007] Furthermore, when a user issues a product quotation task, the permission management agent first verifies the user's identity and access rights, and then the user interaction and demand analysis agent in the user interaction layer converts the user's natural language requirements into structured instructions; the basic data and special expense agent in the data processing layer pulls data from the external system through cross-system call agents, and the price calculation agent integrates all relevant data through multi-objective optimization to generate candidate quotations; finally, the human-computer collaboration and intelligent decision-making assistance agent in the decision-making layer displays all candidate solutions in a visual manner, and experts fine-tune and select the best quotation online. The entire process is ensured by audit logs to ensure the security and traceability of the process.

[0008] Furthermore, the above platform includes the following structures and functions:

[0009] The user interaction and demand analysis agent deployed in the user interaction layer of the platform can segment, denoise and correct spelling of user texts, transcribe user voices in real time through the ASR engine, reserve an extended interface for uploading and parsing images and tables, and support multi-source input such as pictures, drawings, and voice to identify quotation targets; core semantic understanding, using the Transformer large model fine-tuned with industry corpus to perform intent classification and slot extraction, and dynamically inject the quotation domain knowledge graph to ensure that the model has rich business context; in dialogue management and status tracking, all extracted slots and context information are recorded through the session ID, and combined with Preset rules and lightweight reinforcement learning strategies are used to evaluate slot completeness, automatically initiate supplementary questions or secondary confirmations, and update the status once the user responds, and determine when to end the interaction. After the end, the system converts the final confirmed slot into standard JSON containing extensible weight fields (such as cost, winning bid, and value preference) through a template mapping engine, and sends it to the data processing layer through the message bus, while returning the task number and confirmation prompt to the front end. Finally, all original inputs, model predictions, clarification conversations, and user responses are recorded and used to regularly fine-tune the question order and slot priority through reinforcement learning strategies to continuously improve the accuracy of information collection and user experience.

[0010] The data processing layer in the platform includes basic data agent, special fee agent, and price calculation agent:

[0011] Furthermore, the basic data agent, as the data center of the quotation system, pulls heterogeneous original data such as BOM lists, material quotas, labor quotas, purchase prices and inventory status from ERP, MPM, PLM and other systems in real time and in batches through "cross-system calling agents", and completes deduplication, field mapping and format conversion at the access layer, and uniformly transforms the data into an internally usable neutral structure; then, with the help of rule engines and anomaly detection algorithms, the key field data is verified and corrected, and each record is marked with a credibility score in dimensions such as data source, update time and historical hit rate to ensure that subsequent queries use high-quality data first; the cleaned entities (such as products, parts, materials, suppliers and process nodes) and their "assembly relationships", "alternative materials", "cost dependencies", "process sequences" and other associations are mapped to the graph database and updated or rebuilt regularly. The data administrator can correct the graph through the visual interface; on this basis, the graph neural network (GNN) performs multi-hop message transmission in the knowledge graph, automatically Identify implicit cost dependencies and supply chain risk points, and predict material price fluctuation trends; when receiving structured query instructions from the user-interactive intelligent agent, the intelligent agent automatically decomposes tasks according to the "product → subsystem → component → part" hierarchy, calls the knowledge graph and cache layer to return the material usage, unit cost and estimated working hours at each level, and performs availability filtering based on the earliest delivery date and minimum inventory level; to ensure performance in high-concurrency scenarios, the system adopts an LRU strategy to preload high-frequency data and regularly performs consistency checks with the back-end data warehouse; at the same time, it provides a RESTful query interface to the outside world and has built-in Prometheus monitoring and self-healing scripts to automatically switch to the backup data source in the event of a failure; finally, through continuous analysis of indicators such as query time, abnormal recall rate and GNN prediction accuracy, the deviation between the actual transaction price and the estimated value is fed back to the model training pipeline, realizing iterative optimization of cleaning rules, graph structure and inference model, so as to provide high-quality, real-time and reliable material information support for the decision-making layer and human-computer collaboration module.

[0012] Furthermore, the special expense agent, as the core module in the quotation system specifically responsible for various types of test fees, tooling fees, after-sales service fees and other special expenses, first maintains a set of structured expense template libraries internally, classifies common test, tooling, after-sales, one-time special expenses and other items according to process and business scenarios, and supports online expansion and customization; when receiving a structured quotation task, the matching engine will automatically filter out the most relevant expense templates based on the task characteristics, and use the lightweight classification model decision tree to quickly exclude inapplicable items; then, the special expense agent calls the XGB trained with historical data The OOST regression prediction model accurately predicts the floating coefficient for each matched expense item. Incorporating upstream and downstream information such as supplier delivery cycles, logistics delays, and process preparation times, the model automatically identifies the coupling paths between expenses, materials, and processes through a dependency graph module. This allows for synchronous adjustment of related special expenses when base costs or process nodes change. During the prediction and coupling process, the agent's built-in threshold monitor compares the real-time calculated value with the historical average. If the deviation exceeds a preset ratio, an anomaly alert is triggered, marking the expense item for review and notifying the expert. After cleaning and calculation, the agent sends the special expense details to the price calculation module via a unified RESTful API, logging all calculation processes and data sources in a log library for subsequent auditing. Furthermore, the special expense agent integrates Prometheus monitoring and health probes to track model response latency, prediction error, and matching hit rate in real time. When monitored metrics decline, it automatically initiates offline analysis or regression retraining. The deviation between the actual expenditure and the predicted value for this execution is fed back to the model training pipeline to continuously optimize the accuracy of the classification and regression models, ensuring high-quality, explainable, and self-updating special expense support for the quotation system in a dynamic environment.

[0013] Furthermore, the price calculation agent first obtains material costs, labor costs, and detailed expense information from the basic data and special expense agents via a message bus. It then eliminates unreliable or expired records based on a "credibility score." It then combines the cleaned cost data with the user-specified profit margin and tax rate to calculate a preliminary quote. It then automatically applies discounts or spillover factors based on project characteristics to fine-tune costs, generating a base quote and flexibly adjusting the pre-price based on contract terms (such as discounts or prepayment ratios). This agent embeds the NSGA-II genetic algorithm, using user preference weights as optimization criteria to rapidly generate a set of candidate quotes. Based on this, it initiates an adversarial game simulator, extracts historical bidding data from external systems, and conducts multiple rounds of bidding drills for each candidate quote in a reinforcement learning environment to derive the winning probability distribution for each proposal. All candidate quotations and their multi-dimensional evaluation indicators will be packaged into a unified data structure and sent to the human-machine collaboration and solution evaluation module through RESTful API and message bus, or directly synchronized to ERP / CRM for approval in single-solution mode; at the same time, the key data and algorithmic decisions of the entire process, including cost sources, profit margin selection, optimization trajectory, winning probability and sensitivity analysis, will be recorded in an auditable log, and the deviation between the prediction and the actual result will be fed back to the training pipeline of the relevant model, thereby realizing the continuous adaptation and self-optimization of the price calculation agent in a dynamic market environment.

[0014] The platform's decision-making layer deploys a human-machine collaborative and intelligent decision-making assistance agent, which integrates AHP, TOPSIS, and fuzzy comprehensive evaluation algorithms. It calculates the hierarchical weight, ideal solution distance, and membership score for each solution, and then merges the outputs of multiple algorithms into a final ranking based on expert preferences. When experts adjust the profit margin or service commitment of a solution on the interface, the system captures these "adjustment decision points" and immediately triggers a small-scale recalculation, displaying the "before and after adjustment" effect. Finally, the system automatically generates a Top-K recommendation list based on the comprehensive score and expert corrections, with recommendation reasons and risk warnings, and sends the final solution to ERP / CRM with one click, calling permission management to complete the final authorization and audit. All inputs, algorithm decisions, expert modifications, and transaction results are packaged into training samples, and the reinforcement learning feedback engine regularly updates the decision algorithm and visual interaction rules, so that the system is closer to market and customer needs in each quotation, truly realizing "intelligent decision-making."

[0015] Furthermore, the cross-system call agent is a unified hub for data interaction between the platform and external systems (such as ERP, MPM, CRM, unified identity authentication, etc.). It consists of eight modules: asynchronous message listening, interface adaptation, security authentication, data conversion, call scheduling, fault-tolerant degradation, result backfilling and status monitoring. Once the business agent publishes a "target system, interface type, request payload" request through the message bus, the asynchronous message listener immediately captures and forwards it to the interface adapter factory, which instantiates the corresponding adapter (RESTful, GraphQL or proprietary RPC) according to the target system type, and uniformly converts the original request into the protocol and format required by the adapter. Before initiating the call, the security authentication module authenticates and verifies the request based on OAuth2.0 or JWT, and generates or verifies the idempotence key in the idempotence checker. Ensure that repeated requests do not cause repeated writes; the call scheduler is responsible for flow control and frequency limiting of external interfaces to ensure stability in high-concurrency scenarios, and automatically triggers the retry strategy and records each retry log when encountering timeout, 5XX or network failure; if the retry still fails, the fault-tolerant degradation module will switch to the backup data source or initiate a manual alarm according to the preset strategy; regardless of success or failure, the result backfill module will format the response data or error information and return it to the original caller through the message bus, along with the call duration, status code and traffic indicators; finally, the status monitoring component continuously collects key indicators such as call latency, error rate, throughput, etc., and pushes them to Prometheus or ELK cluster, providing upper-level intelligent agents with real-time visual monitoring and automated self-healing trigger conditions, thereby achieving efficient, reliable and traceable cross-system data interaction while ensuring security and compliance.

[0016] Furthermore, the permission management agent is a core component of the entire quotation system to ensure data security, control access rights and ensure operational specifications. The agent integrates the traditional identity authentication mechanism and dynamic context perception, data hierarchical protection and multi-role responsibility isolation to adapt to business changes. The agent adopts multi-factor authentication and single sign-on technology (SSO), and encrypts and verifies user identities and manages lifecycles based on OAuth2.0 and JWT mechanisms to ensure the legitimacy of interface access; the permission allocation integrates role-based access control (RBAC) and attribute-based access control (ABAC) to achieve responsibility isolation between various core roles in the platform, and dynamically adjusts access rights based on the visitor's behavior records, geographic location, time period, device environment and task context. Every piece of data in the system is classified by the data protection engine, marking it as public, internal, restricted, or confidential. The engine then determines access permissions during access and automatically desensitizes or encrypts highly sensitive data. All sensitive operations, permission changes, and cross-system access are fully logged, with this information pushed to the SIEM system in real time for behavioral analysis and compliance audits, helping to identify potential risk paths. Furthermore, the system features a security policy engine that uses pre-set policies to granularly control access rights down to the field level. It also supports policy versioning and rollback mechanisms, ensuring flexible and controllable permission adjustments. To address dynamic risks, the risk early warning system uses AI models to monitor permission usage in real time. If anomalous access patterns are detected (such as repeated logins from different locations or frequent unauthorized operations), the system immediately triggers response mechanisms, including account freezing, forced logout, and multi-level approval. Alerts are also sent to administrators, and the events are simultaneously recorded in the system's security event database for subsequent optimization and training. The permission management agent operates throughout the entire lifecycle of the quotation system, providing robust security and compliance support through a comprehensive, multi-dimensional, and intelligent permission management mechanism.

[0017] Based on the above content, the advantages of the present invention are: achieving fast and accurate quotations for enterprise products, controlling quotation risks through multi-scheme comparison and risk warning, relying on dynamic permissions and audit mechanisms to comprehensively ensure the compliance and security of the quotation process, and continuously optimizing the pricing model through continuous feedback and self-learning, ultimately forming a highly autonomous and sustainably evolving intelligent quotation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the composition of the enterprise product intelligent quotation collaborative decision-making platform in Example 1.

[0019] Figure 2 This is a schematic diagram of the business process of joint execution by intelligent agents in Example 3. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] Example 1

[0022] One embodiment of the present invention discloses an enterprise product intelligent quotation collaborative decision-making platform, the platform composition diagram is as follows Figure 1 As shown in the figure, the platform architecture comprises a user interaction layer, a data processing layer, and a decision-making layer. The user interaction layer deploys a user interaction and demand analysis agent; the data processing layer deploys a basic data agent, a special expense agent, and a price calculation agent; the decision-making layer deploys a human-machine collaboration and intelligent decision-making assistance agent; and the cross-system call agent and permission management agent are deployed within the platform and can be flexibly called upon by other agents based on business needs. Enterprise product quotation tasks are achieved through the collaborative efforts of these agents within the platform.

[0023] Example 2: When a user initiates a product quotation task on the platform:

[0024] In this embodiment, after a user initiates a quotation request for a company's product on the platform, the authority management agent first verifies the user's identity and access rights. Then, the user interaction and demand analysis agent converts the user's natural language requirements into structured task instructions through multiple rounds of dialogue and pushes them to the data processing layer. The basic data agent, with the assistance of the cross-system call agent, obtains BOM, labor quotas and purchase prices from ERP / MPM and other systems in real time, performs data cleaning and knowledge graph matching, and then outputs material and labor cost details. The special expense agent intelligently matches and calculates expense items such as test fees and after-sales service fees, and the price calculation agent concludes. The preliminary draft quotation is generated by combining costs, special expenses and market adjustment factors and all operations are recorded in the log. Then, the human-machine collaboration and intelligent decision-making assistance agent presents the draft quotation to the expert in the form of a visual report. After the expert fine-tunes the key parameters online, the agent recalculates in real time and can perform multi-objective evaluation and ranking of multiple groups of parallel generated solutions based on AHP, TOPSIS or virtual simulation models, and finally select the optimal solution. The system then synchronizes the confirmed quotation back to ERP / CRM through cross-system call agents and pushes it to the customer. In the whole process, the authority management agent monitors and records the operation trajectory throughout to ensure safety, compliance and auditability.

[0025] In Example 3, the user adds special requirements and initiates a product quotation task with the three dimensions of "cost control, bid winning probability, and customer value" as the focus. The intelligent agent jointly executes the business process as follows Figure 2 As shown:

[0026] In this embodiment, the business process is as follows:

[0027] 1.Task initiation and permission management

[0028] After a user submits a quotation request for "500 units of X-200 model products including five years of maintenance" on the platform, the permission management agent immediately intercepts and performs multi-factor authentication (MFA) and single sign-on (SSO) verification, verifies the user's access rights to quotation data, contract terms and historical orders based on RBAC+ABAC rules, and writes the operation into an unalterable audit log.

[0029] 2. Requirements Capture and Scenario Completion

[0030] The user interaction and demand analysis agent uses ASR or text preprocessing to parse slots such as "Model X-200", "Quantity 500", "Five-year Maintenance", and "Delivery Period". It combines lightweight reinforcement learning strategies and business rules to actively ask users about their emphasis on the three dimensions of "cost control, probability of winning the bid, and customer value", and encapsulates all slots and initial weights into standard JSON instructions.

[0031] 3. External data preheating

[0032] The JSON instruction is issued in parallel by the cross-system calling agent: pulling BOM, labor time quota, purchase price and inventory from ERP / MPM; obtaining customer transaction history and competitive bid price from CRM / bidding platform; synchronizing the latest security policies and user behavior logs to the unified authentication system, and aggregating these data into the cache.

[0033] 4. Basic Data Integration and Graph Reasoning

[0034] The basic data agent performs ETL cleaning, field verification and credibility scoring on the raw data in the preheated cache, maps products, components, materials and processes to the graph database, and loads the graph neural network model to perform multi-hop dependency completion and single supplier risk identification, and finally derives the material cost and labor cost details of each level of components.

[0035] 5. Special expense matching and forecasting

[0036] The special expense agent calls the classification engine matching template based on the keywords "test fees, tooling fees, and five-year maintenance" in the task requirements, and then uses the regression model trained with historical data combined with the supplier delivery time and logistics delay prediction floating coefficient to automatically mark the cost items with abnormal deviations as "pending review."

[0037] 6. Cost consolidation and verification

[0038] The price calculation agent combines the basic cost and special expenses into the total cost, and uses a cross-validation algorithm to check the consistency of key cost items, eliminate abnormal feedback data, and form a "net cost baseline."

[0039] 7. Profit, Tax and Contract Terms Markup

[0040] The price calculation agent reads the target profit margin and regional tax rate in the task instructions as a baseline cost plus, dynamically adjusts the terms based on contract discounts, advance payment conditions, etc., and generates a preliminary quotation including profit and taxes.

[0041] 8. Market Simulation and Rivalry

[0042] The cross-system call agent pulls the latest industry price index and historical bidding quotations. The price calculation agent has a built-in opponent simulation sub-module that builds a bidding strategy based on reinforcement learning. After multiple rounds of adversarial simulation, the winning probability distribution of each pre-quotation is output.

[0043] 9. Multi-objective optimization and sensitivity analysis

[0044] The price calculation agent inputs "cost fit, winning probability, customer value" and user weights into the improved NSGA-II optimization engine to generate Pareto frontier candidate quotations, perform local perturbation tests on each candidate, and output "cost sensitivity radar" and "price-winning rate curve" reports.

[0045] 10. Visual Decision Making and Expert Fine-tuning

[0046] The human-machine collaboration and intelligent decision-making auxiliary agent presents candidate solutions in a three-dimensional visualization dashboard of "cost risk-winning bid income-value premium", allowing experts to drag weight sliders or directly adjust parameters (such as profit margin, service commitment). The system recalculates and updates the view in real time, and records the effects of each adjustment and the "before and after comparison".

[0047] 11. Final confirmation and system implementation

[0048] After the expert confirms the final quotation, the cross-system call agent will synchronize the results back to ERP / CRM, triggering the contract approval process; the permission management agent performs real-time authentication and auditing on all data reading and writing, parameter adjustment and approval operations to ensure security and compliance.

[0049] 12. Closed-loop feedback and continuous iteration

[0050] After the project is executed, the reinforcement learning feedback engine compares the "predicted vs. actual" winning rate, cost and customer satisfaction, and feeds the deviation data back to the basic data cleaning, special expense forecasting, opponent simulation and optimization models, automatically triggering retraining and strategy iteration of each module on a regular basis to achieve continuous adaptation and lean optimization of the system.

[0051] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An enterprise product intelligent quotation collaborative decision-making platform, characterized by: The platform architecture includes a user interaction layer, a data processing layer, and a decision-making layer. The functions of each layer in the platform are realized through the collaboration of multiple intelligent agents, including: The user interaction and demand analysis agents are deployed in the user interaction layer; the basic data agents, special expense agents, and price calculation agents are deployed in the data processing layer; the human-machine collaboration and intelligent decision-making assistance agents are deployed in the decision-making layer; the cross-system call agent and permission management agent are deployed in the platform and can be flexibly called by other agents according to business needs. The platform workflow is as follows: When a user posts a product quotation task, the permission management agent first verifies the user's identity and access rights. Then, the user interaction and demand analysis agent in the user interaction layer converts the user's natural language requirements into structured instructions. After receiving the instructions, the basic data and special expense agent in the data processing layer pulls data from the external system through a cross-system call agent. The price calculation agent integrates and calculates relevant data through multi-objective optimization to generate candidate quotations. Finally, the human-computer collaboration and intelligent decision-making assistance agent in the decision-making layer displays all candidate solutions in a visual manner. Experts fine-tune and select the best quotation online. The entire process information is recorded by the audit log.

2. The enterprise product intelligent quotation collaborative decision-making platform according to claim 1 is characterized in that: The user interaction and demand parsing agent in the user interaction layer supports users to request multi-source input in the form of text, voice, and pictures, uses the industry-fine-tuned deep learning Transformer architecture model based on the self-attention mechanism to classify user intentions and extract slots, and dynamically injects the quotation domain knowledge graph to give the model a multi-business context; automatically identifies missing information and initiates supplementary questions through dialogue status tracking with users and lightweight reinforcement learning strategies until all slots are determined to be complete and the process is completed; finally, the structured results are mapped into standard JSON with weight fields, sent to the data layer via the message bus, and the entire interaction log is recorded for subsequent optimization.

3. The enterprise product intelligent quotation collaborative decision-making platform according to claim 1 is characterized in that: The basic data agent in the data processing layer calls the product-related heterogeneous raw data in the aggregated enterprise resource planning system (ERP), manufacturing process management system (MPM), and product lifecycle management system (PLM) through the "cross-system calling agent". After cleaning, verifying and scoring the multi-dimensional credibility of the data, the product, component, material and their "assembly-substitution-cost dependency" relationship are mapped to the knowledge graph, and the graph neural network is used to mine implicit cost dependency and supply chain risks and predict price trends. When receiving a structured query, the task is automatically decomposed according to the product → subsystem → component → part level, and the material usage, unit cost and working hour report are returned. The system supports standardized query interfaces and online monitoring, and feeds back the actual transaction and forecast deviation to the model training pipeline, continuously optimizing the graph structure and data cleaning and reasoning capabilities.

4. The enterprise product intelligent quotation collaborative decision-making platform according to claim 1 is characterized in that: The data processing layer contains a special expense agent that categorizes product testing, tooling, after-sales, and one-time special expenses according to process flows and business scenarios, and maintains an expandable library of structured expense templates internally. After receiving a quotation task, it quickly matches relevant templates through a decision tree, calls the XGBoost regression model trained with historical data to predict the floating coefficient of each expense, and combines upstream and downstream information such as supplier delivery periods, logistics delays, and process preparation times. It automatically identifies the coupling paths between expenses and materials and processes through dependency graphs, and when basic costs or process nodes change, the relevant special expenses can be adjusted synchronously. During the prediction and coupling process, the agent has a built-in threshold monitor. Once the real-time threshold detects a large deviation from the historical mean, it triggers an alert and marks it for review. The agent sends the cleaned and verified special expense details to the price calculation module through a RESTful API, and records all process data in a log library. The system integrates the Prometheus monitoring system and health probes to record the deviation between model performance and actual expenditure, so as to provide continuous feedback and optimize the classification and regression models.

5. The enterprise product intelligent quotation collaborative decision-making platform according to claim 1 is characterized in that: The price calculation agent in the data processing layer obtains and filters basic cost data through the message bus, merges and processes the data after cleaning to generate a preliminary quotation; embeds the NSGA-II genetic algorithm, uses user preference weights as optimization criteria, and quickly generates candidate quotations; captures historical bidding data from external systems, and evaluates the probability of winning the bid by simulating competing strategies through reinforcement learning; all quotation data is packaged and sent to the decision module via the RESTful API and message bus. In the single-scheme mode, it is directly synchronized to the enterprise resource planning system (ERP) / customer relationship management system (CRM) to initiate approval. During this period, process data is recorded in the audit system, and the deviation between prediction and actual is fed back to the model training pipeline to achieve continuous adaptive optimization of the price calculation agent.

6. The enterprise product intelligent quotation collaborative decision-making platform according to claim 1 is characterized in that: The decision-making layer uses human-machine collaboration and an intelligent decision-making support agent, which integrates the Analytic Hierarchy Process (AHP), the Top-of-Severe Solutions (TOPSIS) method, and fuzzy comprehensive evaluation to score and rank pricing proposals. When experts adjust proposals online, the system captures the adjustment points and recalculates and compares the results in real time. Finally, based on the comprehensive scores and expert corrections, a Top-K recommendation list is generated, along with the recommendation reasons and risk warnings. This list is then distributed to the Enterprise Resource Planning (ERP) and Customer Relationship Management (CRM) systems with a single click, completing authority audits. The input, decision-making and feedback of the entire process are stored as training samples, and reinforcement learning is used to regularly optimize the algorithm and interaction rules to ensure that the output results are in line with market and customer needs.

7. The enterprise product intelligent quotation collaborative decision-making platform according to claim 1 is characterized in that: The platform calls an intelligent agent across systems, which receives data retrieval instructions from each business module through the message bus. Through asynchronous listeners, interface adapters, security authentication, idempotence verification and flow control scheduling, the agent automatically converts the request into RESTful / GraphQL / RPC format and sends it to the external system. In case of failure, it performs retries or downgrades to switch to the backup data source, and finally formats the response result and returns it to the original caller through the message bus. Through online monitoring, it collects delay and error rate indicators to provide data interaction services for the upper-level intelligent agent.

8. The enterprise product intelligent quotation collaborative decision-making platform according to claim 1 is characterized in that: The platform's internal permission management agent uses multi-factor authentication and single sign-on (SSO) technology to encrypt and verify user identities and manage their lifecycles based on the OAuth 2.0 authorization framework and JWT (JSON Web Token) mechanism. Permission allocation integrates role-based access control (RBAC) and attribute-based access control (ABAC) to achieve duty isolation between various core roles in the platform, and dynamically adjusts access rights based on the visitor's behavior records, geographic location, time period, device environment, and task context. Data is classified into "public-internal-restricted-confidential" categories, and highly sensitive data is encrypted and desensitized for transmission. All sensitive operations, permission changes, and cross-system access are recorded in the SIEM security management system audit log. At the same time, the security policy engine uses preset policies to fine-tune access rights to the field level and supports policy version management and rollback mechanisms. The risk warning system monitors permission usage behavior in real time based on AI models, and immediately triggers a response mechanism when abnormal access patterns are detected.

Images