Method And Apparatus for Generating Real-Time Personalized Content Based on Large Language Model and AI Agent

Abstract

A method and apparatus for generating real-time customized content based on a large-scale language model and an AI agent are disclosed. The present embodiment provides a method and apparatus for generating real-time customized content based on a large-scale language model and an AI agent, which generate and send real-time customized content across industries including newsletters, foreign language education, e-commerce marketing, financial reports, and media content, by utilizing artificial intelligence, specifically a large-scale language model (LLM) and an AI agent, to generate, edit, and send content.

Description

Technology Field

[0001] One embodiment of the present invention relates to a method and apparatus for generating real-time customized content based on a large-scale language model and an AI agent. Background Technology

[0002] The following description merely provides background information related to the present embodiment and does not constitute prior art.

[0003] Traditional content delivery systems rely on periodic, standardized, and manual editing methods, lacking personalization and real-time capabilities. In the field of education (EduTech), the provision of customized content tailored to learners' levels is limited.

[0004] Therefore, in the field of marketing, it is difficult to ensure real-time product recommendations and copy delivery tailored to individual consumers. The aforementioned issues lead to reduced operational efficiency, decreased user satisfaction, and lower retention. The problem to be solved

[0005] The present embodiment aims to provide a method and apparatus for generating real-time customized content based on a large-scale language model and an AI agent, which generate and send real-time customized content across various industries including newsletters, foreign language education, e-commerce marketing, financial reports, and media content, by utilizing artificial intelligence, specifically a large-scale language model (LLM) and an AI agent, to generate, edit, and send content. means of solving the problem

[0006] According to one aspect of the present embodiment, a customized content generation device is provided, comprising: a data collection unit for collecting external data; an LLM analysis unit for generating an LLM analysis result by performing an LLM analysis on the external data; a user profiling unit for generating a user profile result by performing a user-specific profile based on the LLM analysis result; a content generation unit for generating and sending personalized content based on the LLM analysis result and the user profile result; and a feedback loop providing unit for optimizing the LLM analysis result based on feedback regarding the personalized content.

[0007] According to another aspect of the present embodiment, a real-time customized content generation method is provided, characterized by comprising: a process in which a data collection unit collects external data; a process in which an LLM analysis unit performs LLM analysis on the external data and generates an LLM analysis result; a process in which a user profiling unit performs user-specific profiling based on the LLM analysis result and generates a user profile result; a process in which a content generation unit generates and sends personalized content based on the LLM analysis result and the user profile result; and a process in which a feedback loop providing unit optimizes the LLM analysis result based on feedback regarding the personalized content. Effects of the invention

[0008] As explained above, according to the present embodiment, by utilizing artificial intelligence, particularly a large-scale language model (LLM) and an AI agent, to generate, edit, and send content, there is an effect of generating and sending real-time customized content across various industries, including newsletters, foreign language education, e-commerce marketing, financial reports, and media content. Brief explanation of the drawing

[0009] FIG. 1 is a diagram showing a real-time customized content generation system according to the present embodiment. FIG. 2 is a drawing showing a customized content generation device according to the present embodiment. FIG. 3 is a diagram illustrating a method for generating real-time customized content according to the present embodiment. Specific details for implementing the invention

[0010] Hereinafter, the present embodiment will be described in detail with reference to the attached drawings.

[0011] FIG. 1 is a diagram showing a real-time customized content generation system according to the present embodiment.

[0012] The real-time customized content generation system according to the present embodiment includes a terminal (110), a network (120), and a customized content generation device (130). The components included in the real-time customized content generation system are not necessarily limited thereto.

[0013] A terminal (110) refers to an electronic device that performs voice or data communication via a network (120) according to the user's key operation.

[0014] The terminal (110) is equipped with a memory for storing a program or protocol for communicating with a customized content creation device (130) via a network (120), a microprocessor for executing the program to perform calculations and control, etc.

[0015] The terminal (110) includes any one of the following electronic devices: a smartphone, a tablet, a laptop, a personal computer (PC), a personal digital assistant (PDA), a portable multimedia player (PMP), a wireless communication terminal, and a media player.

[0016] The terminal (110) is a various device equipped with a communication device, such as a communication modem, for communicating with various devices or wired / wireless networks, a memory for storing various programs and data, and a microprocessor for executing programs to perform calculations and control. According to at least one embodiment, the memory may be a computer-readable recording / storage medium such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, an optical disk, a magnetic disk, or a Solid State Disk (SSD). According to at least one embodiment, the microprocessor may be programmed to selectively perform one or more operations and functions described in the specification. According to at least one embodiment, the microprocessor may be implemented as hardware, such as an Application Specific Integrated Circuit (ASIC), wholly or partially.

[0017] The terminal (110) connects to the customized content creation device (130) via the network (120) and receives customized content.

[0018] The network (120) refers to a network capable of transmitting and receiving data via the Internet protocol using various wired and wireless communication technologies, such as the Internet network, intranet network, mobile communication network, and satellite communication network. The network (120) relays data between the terminal (110) and the customized content creation device (130).

[0019] The customized content generation device (130) includes hardware modules identical to those of a conventional web server, WAP server, or network server. The customized content generation device (130) generally communicates with an unspecified number of clients or other servers via an open computer network such as the Internet. The customized content generation device (130) refers to a computer system or computer software (web server program) that derives and provides work results in response to work execution requests from clients or other web servers. In addition to the aforementioned web server program, the customized content generation device (130) includes a series of application programs operating on the web server or various databases built within the device.

[0020] The customized content generation device (130) provides real-time data collection and automatic content summarization, translation, and processing based on LLM. The customized content generation device (130) performs AI agent-based user profiling and sending optimization. The customized content generation device (130) supports the creation of industry-specific content such as foreign language education, e-commerce, and finance. The customized content generation device (130) sends via multiple channels (email, application, platform) and reflects user feedback. The customized content generation device (130) secures general-purpose content infrastructure technology that is scalable across industries in the long term.

[0021] The customized content generation device (130) improves the level of personalization and real-time capabilities compared to a general sending system. The customized content generation device (130) maximizes operational efficiency by reducing manual work by operators. The customized content generation device (130) enables utilization in various markets by utilizing industry-specific scalability. The customized content generation device (130) provides standardized platform technology for content infrastructure in the long term.

[0022] The customized content generation device (130) generates and automatically sends personalized newsletters that include real-time economic and financial information curation. The customized content generation device (130) generates and automatically sends personalized learning materials and quizzes tailored to the level and goals of the learner for foreign language education. The customized content generation device (130) automatically generates marketing copy based on product information and recommends and sends personalized e-commerce. The customized content generation device (130) generates and automatically sends personalized time market data analysis and automatic reports. The customized content generation device (130) generates and automatically sends personalized media / entertainment that includes drama / variety show related curation and fan-tailored notifications.

[0023] The customized content generation device (130) automatically generates personalized investment / financial reports. The customized content generation device (130) generates personalized investment strategy reports based on real-time market data. The customized content generation device (130) provides personalized entertainment / media curation. The customized content generation device (130) provides personalized content based on personalized fandom and viewer reactions. The customized content generation device (130) sends personalized healthcare information. The customized content generation device (130) generates real-time management reports based on patient health data.

[0024] The customized content generation device (130) generates customized content based on real-time data using a large-scale language model and an AI agent, and then sends it to a terminal (110). The customized content generation device (130) generates customized content that includes learning materials with difficulty levels adjusted according to the level of a foreign language learner. The customized content generation device (130) generates customized content that includes e-commerce product descriptions and personalized marketing copy. The customized content generation device (130) generates customized content that includes real-time financial and investment reports. The customized content generation device (130) sends customized content through multiple channels using a sending module that includes email, applications, messengers, and online platforms.

[0025] FIG. 2 is a drawing showing a customized content generation device according to the present embodiment.

[0026] The customized content generation device (130) according to the present embodiment includes a data collection unit (210), an LLM analysis unit (220), a user profiling unit (230), a content generation unit (240), and a feedback loop providing unit (250). The components included in the customized content generation device (130) are not necessarily limited thereto.

[0027] Each component included in the customized content creation device (130) can be connected to a communication path connecting a software module or a hardware module inside the device and operate organically with one another. These components communicate using one or more communication buses or signal lines.

[0028] Each component of the customized content creation device (130) illustrated in FIG. 2 represents a unit that processes at least one function or operation, and can be implemented as a software module, a hardware module, or a combination of software and hardware.

[0029] The data collection unit (210) collects data by performing real-time crawling on external news, educational materials, product information, and media content. The data collection unit (210) dynamically collects multi-source data for data processing and collection. The data collection unit (210) performs real-time priority queue-based data processing to simultaneously collect news, social media, e-commerce APIs, and educational platform materials. The data collection unit (210) performs noise removal and reliability-based content filtering to automatically remove false information and duplicate information and calculate reliability scores.

[0030] The data collection unit (210) according to the present embodiment collects external data. The data collection unit (210) collects external data by performing real-time crawling on external news, educational materials, product information, and media content.

[0031] The LLM analysis unit (220) performs summarization, translation, level-specific difficulty adjustment, and copywriting generation for LLM analysis and processing. The LLM analysis unit (220) automatically generates multi-level learning content for LLM content processing. The LLM analysis unit (220) transforms the same text into A1~C2 (beginner~advanced) difficulty levels. The LLM analysis unit (220) performs real-time multilingual translation and cultural context reflection. The LLM analysis unit (220) performs country-specific context / tone reflection rather than simple translation. The LLM analysis unit (220) automatically optimizes product descriptions / marketing copy. The LLM analysis unit (220) transforms copywriting based on product characteristics and user profiles.

[0032] The LLM analysis unit (220) according to the present embodiment generates an LLM analysis result that performs LLM analysis on external data. The LLM analysis unit (220) performs an LLM analysis that performs at least one of summarizing, translating, level-based difficulty adjustment, and copywriting on the external data.

[0033] The LLM analysis unit (220) performs LLM analysis through multiple pipelines. The LLM analysis unit (220) performs preprocessing that involves external data cleaning and semantic unit partitioning. The LLM analysis unit (220) configures prompts for generating automatic templates based on summarization, translation, difficulty, and copywriting purposes. The LLM analysis unit (220) performs LLM analysis that performs actual conversion using LLM. The LLM analysis unit (220) performs postprocessing that corrects the stylistic and structural consistency of the output. The LLM analysis unit (220) performs output and feedback that continuously improves by transmitting to the user profiling unit and reflecting feedback.

[0034] The LLM analysis unit (220) performs input data normalization and preprocessing. When the LLM analysis unit (220) receives external data from the data collection unit (210), it first converts the external data to be suitable for the LLM input format. The LLM analysis unit (220) performs a preprocessing process for the external data. The LLM analysis unit (220) performs text cleaning on the external data, such as removing HTML tags, removing duplicate content, and automatically correcting missing or erroneous sentences. The LLM analysis unit (220) performs semantic unit partitioning on the external data, which reclassifies data such as news, product information, and educational materials into paragraph units or topic units. The LLM analysis unit (220) performs metadata attachment on the external data, such as category, language, extraction time, and data reliability score. The LLM analysis unit (220) transmits the preprocessed input text to the LLM in the next step.

[0035] The LLM analysis unit (220) configures LLM analysis prompts. The LLM analysis unit (220) generates different prompt templates depending on the type of analysis to be performed (summary, translation, difficulty adjustment, copywriting).

[0036] The LLM analysis unit (220) has the following prompt configuration method.

[0037] The LLM analysis unit (220) configures a summary prompt template. The LLM analysis unit (220) automatically selects a prompt according to the purpose, such as extracting key sentences, structuring summaries, and generating highlights.

[0038] The LLM analysis unit (220) generates a translation prompt template. The LLM analysis unit (220) includes country-specific cultural context, tone (formal, friendly), and domain terminology in the translation prompt template. The LLM analysis unit (220) “reflects cultural context / tone rather than simple translation” in the translation prompt template.

[0039] The LLM analysis unit (220) configures a difficulty adjustment prompt. The LLM analysis unit (220) configures a difficulty adjustment prompt based on CEFR (A1~C2) or a difficulty index. The LLM analysis unit (220) automatically adjusts sentence length, vocabulary difficulty, grammatical structure, etc. within the difficulty adjustment prompt. The LLM analysis unit (220) reflects the transformation of the same text into A1~C2 difficulty within the difficulty adjustment prompt.

[0040] The LLM analysis unit (220) configures a copywriting prompt. The LLM analysis unit (220) reflects product characteristics and user profiles (interests and purchase history) in the copywriting prompt. The LLM analysis unit (220) reflects modifications to the copywriting prompt based on product characteristics and user profiles.

[0041] The LLM analysis unit (220) performs actual analysis using an LLM model. The LLM analysis unit (220) inputs the generated prompt and preprocessed text into an LLM engine (based on an internal or external API) to perform summarization, translation, level-based difficulty adjustment, and copywriting.

[0042] The LLM analysis unit (220) performs summarization using an LLM model. The LLM analysis unit (220) summarizes preprocessed data by performing key sentence extraction, semantic compression, and length adjustment for each user. The LLM analysis unit (220) automatically selects a summary type for a specific purpose, such as a newsletter or financial report.

[0043] The LLM analysis unit (220) performs translation using an LLM model. The LLM analysis unit (220) performs multilingual translation. The LLM analysis unit (220) reflects cultural context (country-specific context, tone) and domain characteristics.

[0044] The LLM analysis unit (220) performs level-by-level difficulty adjustment using an LLM model. The LLM analysis unit (220) performs readjustment of sentence length, vocabulary, and grammatical complexity according to CEFR levels. The LLM analysis unit (220) automatically generates text suitable for learner levels in the field of education.

[0045] The LLM analysis unit (220) performs copywriting using an LLM model. The LLM analysis unit (220) performs CTA (Call to Action), emotional expression, and product USP emphasis. The LLM analysis unit (220) performs tone adjustment based on user profiles.

[0046] The LLM analysis unit (220) performs result quality correction (post-processing). The LLM analysis unit (220) performs post-processing to ensure structural and stylistic consistency of the LLM output. For post-processing, the LLM analysis unit (220) performs verification of sentence naturalness, application of prohibited words or brand tone guides, removal of traces of machine translation, and removal of redundant summaries or similar expressions to improve the completeness of the final LLM analysis result.

[0047] The LLM analysis unit (220) outputs the analysis results and transmits them to the user profiling unit (230). The LLM analysis unit (220) transmits the LLM analysis results, converted into a structured form (JSON, paragraph, category form, etc.), to the user profiling unit (230). The LLM analysis unit (220) enables the user profiling unit (230) to generate personalized recommendation materials by combining learning levels, purchase history, and content consumption patterns based on the LLM analysis results.

[0048] The user profiling unit (230) performs personalization based on learning level, purchase history, and content consumption patterns. The user profiling unit (230) provides a real-time recommendation algorithm based on user behavior data. The user profiling unit (230) adjusts the dynamics of sent content according to clicks, purchases, and learning history for each user. The user profiling unit (230) determines the optimal sending timing for each time zone and device. For example, the user profiling unit (230) performs automatic sending of learning materials in the morning and marketing content in the evening.

[0049] The user profiling unit (230) according to the present embodiment generates a user profile result by performing a user-specific profile based on the LLM analysis result. The user profiling unit (230) performs a user-specific profile including at least one of a learning level, purchase history, and content consumption pattern based on the LLM analysis result.

[0050] The user profiling unit (230) performs user-specific profiling in multiple stages. The user profiling unit (230) collects LLM analysis results and user behavior data. The user profiling unit (230) extracts features of learning level, purchase history, and content consumption pattern. The user profiling unit (230) generates user embeddings and performs clustering analysis. The user profiling unit (230) generates personalization rules based on the LLM analysis results. The user profiling unit (230) transmits the individual profile results to the content creation unit (240). The user profiling unit (230) implements a function of “performing user-specific profiling including at least one of learning level, purchase history, and content consumption pattern based on LLM analysis results.”

[0051] The user profiling unit (230) performs input data aggregation (integration of LLM analysis results and user behavior data). The user profiling unit (230) collects the LLM analysis results, user behavior data, and internal system data together and performs integrated analysis.

[0052] The user profiling unit (230) collects LLM analysis results (provided by the LLM analysis unit). The user profiling unit (230) collects text difficulty (distinguishing between A1 and C2 levels), category (learning / news / marketing / finance, etc.), major keywords and topics, product characteristics, or educational goals.

[0053] The user profiling unit (230) collects user behavior data. The user profiling unit (230) collects learning activities including correct answer rates, problem-solving time, and perceived response by difficulty level. The user profiling unit (230) collects purchase history including purchased items, interest categories, and repurchase cycles. The user profiling unit (230) collects content consumption patterns including viewing time, click-through rate, scroll patterns, and reading completion rate.

[0054] The user profiling unit (230) collects internal system data. The user profiling unit (230) collects user device information, connection time zones, behavior flow between sessions, etc.

[0055] The user profiling unit (230) integrates the LLM analysis results, user behavior data, and internal system data to generate a user-specific profile input vector.

[0056] The user profiling unit (230) extracts features according to the purpose of profiling. The user profiling unit (230) extracts key features according to the purpose.

[0057] The user profiling unit (230) extracts the learning level (feature).

[0058] The user profiling unit (230) calculates the learning level of individual users by combining the 'text difficulty (A1~C2)' and the user's learning response data from the LLM analysis results. The user profiling unit (230) calculates the learning level based on difficulty. The user profiling unit (230) determines the user as a candidate for a higher level if the accuracy rate increases when consuming content of a specific difficulty level, and determines the user as a candidate for a lower level if the dwell time is excessively long and the error rate increases. The user profiling unit (230) automatically corrects the level by analyzing changes in learning patterns (speed and accuracy) when the same topic is studied repeatedly. For example, if the accuracy rate of A2 difficulty content for an English learner is 80% or higher and the accuracy rate of B1 content is 40% or lower, the user profiling unit (230) profiles the user as "A2.5 level" or "between A2 and B1."

[0059] The user profiling unit (230) extracts purchase history (feature). The user profiling unit (230) combines purchase data with LLM analysis to determine the user's product preference.

[0060] The user profiling unit (230) analyzes purchase history. The user profiling unit (230) combines purchase data with LLM analysis to analyze purchase history including price range trends (low / mid / premium), category preferences (fashion > home appliances > food, etc.), and brand loyalty. The user profiling unit (230) extracts interest elements based on LLM keywords extracted from product descriptions (“eco-friendly,” “gaming,” “premium material,” etc.).

[0061] The user profiling unit (230) generates profiling results. The user profiling unit (230) generates profiling results including “customers who prefer premium electronic products,” “customers who are sensitive to eco-friendly products,” and “customers with strong repeat purchase patterns (subscription-based recommendations possible).”

[0062] The user profiling unit (230) extracts content consumption patterns (features). The user profiling unit (230) models how users react to various forms of content. The user profiling unit (230) analyzes consumption patterns based on viewing time, click-through rate (CTR), scroll depth and dwell time, reading breakpoint (mid-break pattern), preferred content by time of day, and consumption trends by device (mobile / PC).

[0063] The user profiling unit (230) generates profiling results. The user profiling unit (230) generates profiling results including: Morning: more clicks on news / learning materials, Afternoon / Night: increased click-through rate on marketing copy, and automatic optimization of sending policies by time of day.

[0064] The user profiling unit (230) performs user feature combination (User Embedding) and clustering. The user profiling unit (230) integrates the previously extracted 'learning level, purchase history, and consumption pattern' into a single user embedding (User Vector). For example, the user profiling unit (230) generates a profile vector [learning level index, category preference, purchase pattern index, consumption pattern index, time zone response pattern, LLM keyword matching score ...]. The user profiling unit (230) can classify the profile vector into profile groups through a clustering algorithm (K-means, DBSCAN, etc.) or rule-based clustering. For example, the user profiling unit (230) can classify the profile vector into forms such as “beginner learner + active consumer + evening click-intensive type,” “advanced learner + specialized news preference type,” and “light consumer + mobile-based content consumption type.”

[0065] The user profiling unit (230) generates personalization rules based on the results of LLM analysis. The user profiling unit (230) automatically generates personalization rules to be used based on the derived profile. For example, if the user profiling unit (230) has a learning level of A2, it prioritizes recommending content with difficulty levels from A2 to B1. If the user profiling unit (230) determines that there is a preference for eco-friendly products, it generates copy that emphasizes the 'eco-friendly keyword' in the product description. If the user profiling unit (230) identifies a user with an increase in CTR in the afternoon, it prioritizes sending during the afternoon time slot.

[0066] The user profiling unit (230) generates final user profile data and transmits it to the content creation unit (240). The user profiling unit (230) generates a final profile that includes user learning level (e.g., B1, beginner / intermediate index, etc.), purchasing propensity data (category preference, price range, brand loyalty, etc.), content consumption patterns (response patterns by time of day, format, topic), and a set of personalization rules (rules applied to learning, marketing, newsletters, etc.).

[0067] The user profiling unit (230) transmits the final profile to the content creation unit (240) so that personalized newsletters / learning materials / copywriting / reports are generated.

[0068] The content creation unit (240) supports multiple formats such as newsletters, learning materials, marketing copy, and financial reports. The content creation unit (240) sends content using email, applications, platform APIs, and messenger-based multi-channels.

[0069] The content creation unit (240) performs sending and channel optimization. The content creation unit (240) performs simultaneous sending via multiple channels and integrated data management. The content creation unit (240) manages email, application push, messenger, and learning platform sending from a single engine. The content creation unit (240) performs automatic recovery and retries for sending failures. The content creation unit (240) responds to server errors and email rejections and performs automatic resending.

[0070] The content generation unit (240) according to the present embodiment generates and sends personalized content based on LLM analysis results and user profile results. The content generation unit (240) generates personalized content including at least one of a newsletter, learning materials, marketing copy, and financial reports based on LLM analysis results and user profile results.

[0071] The content creation unit (240) transmits personalized content to the terminal (110) using email, an application, a platform API, and a messenger-based multi-channel.

[0072] The content generation unit (240) generates personalized content through multiple procedures. The content generation unit (240) determines the content type and selects a template. The content generation unit (240) generates content based on LLM analysis results (summary / translation / difficulty / copy / keywords) and user profile-based personalization rules (learning level, purchase history, consumption pattern). The content generation unit (240) performs LLM-based final document regeneration and post-processing. The content generation unit (240) prepares for multi-channel distribution after formatting and packaging. The content generation unit (240) automatically generates content different for each user, such as newsletters, learning materials, marketing copy, and financial reports.

[0073] The content creation unit (240) determines the purpose of content creation and selects a template.

[0074] The content creation unit (240) first determines the type of content required by the user based on the user profile. If the user is a learner, the content creation unit (240) prioritizes the creation of learning materials (e.g., English reading passages by difficulty level, quizzes). If the user is an e-commerce user, the content creation unit (240) determines marketing copy or product recommendation newsletters as the required content type. If the user is an investment-interested user, the content creation unit (240) determines financial reports or market data summaries as the required content type. If the user is a media consumer, the content creation unit (240) determines news / trend-based newsletters as the required content type.

[0075] The content creation unit (240) provides a template loading method including a newsletter template, a learning material question template (by difficulty level A1~C2), a marketing copy template (CTA, one-line USP composition, etc.), and a financial report template (summary, portfolio, chart explanation, etc.).

[0076] The content creation unit (240) reflects the LLM analysis results.

[0077] The content creation unit (240) utilizes summarized key content provided from the LLM analysis results, translated / multilingual writing style, text with adjusted difficulty (A1~C2), keywords by category, descriptions of product characteristics or financial indicators, and a tone & writing style that reflects cultural context.

[0078] The content generation unit (240) automatically inserts "summarized article + key points + related links" when generating a newsletter. When generating learning materials, the content generation unit (240) automatically generates A2 difficulty text and related questions generated by LLM. When generating marketing copy, the content generation unit (240) generates text centered on the USP (Key Selling Point) extracted by LLM. When generating financial reports, the content generation unit (240) reconstructs LLM's article summary + indicator interpretation with a unified writing style.

[0079] The content creation unit (240) applies user profile-based personalization rules.

[0080] The content creation unit (240) converts the LLM analysis results into “personalized content” by reflecting the user profile results.

[0081] The content generation unit (240) reflects the learning level. The content generation unit (240) provides A2 level text and easy vocabulary explanations to A2 level learners. The content generation unit (240) provides C1 level learners with advanced sentences and advanced problems. For example, the content generation unit (240) provides automatic adjustment of sentence structure, vocabulary level, and problem difficulty for “English learning content”.

[0082] The content generation unit (240) reflects purchase history. The content generation unit (240) generates copy focused on high-end new products for premium electronic product buyers. The content generation unit (240) generates phrases emphasizing eco-friendly keywords for eco-friendly product preferences. The content generation unit (240) automatically recommends a skincare / beauty new product newsletter to recent beauty category buyers. The content generation unit (240) causes the tone of the copy phrases to change according to user preferences.

[0083] The content generation unit (240) reflects content consumption patterns. If the user prefers short text, the content generation unit (240) organizes it around summaries. The content generation unit (240) provides analytical and long-form content to users who read for a long time. The content generation unit (240) applies a tone (soft writing style) for nighttime delivery to users with high nighttime click-through rates. For example, if the content generation unit (240) determines that “this user stayed on financial data analysis content for a long time,” it generates content by strengthening the chart explanations in financial reports.

[0084] The content creation unit (240) performs LLM-based regeneration and post-processing.

[0085] For example, the content creation unit (240) can provide the LLM with a prompt in the format including 'User Level: B1, Preferred Category: Eco-friendly / Home Appliances, Content Type: Newsletter, Tone: Concise and Friendly, Elements to Include: 3 Latest Eco-friendly Home Appliance Trends and Summary' based on user-customized parameters. The content creation unit (240) uses the LLM to create personalized final customized content based on the prompt in the format. After creating personalized content, the content creation unit (240) performs post-processing such as filtering prohibited words and sensitive expressions, unifying the writing style, removing grammatical errors and repetitions, and applying a final style that conforms to the brand guide.

[0086] The content creation unit (240) performs individual content packaging and preparation for sending.

[0087] The content generation unit (240) converts the generated customized content to fit the sending format. The content generation unit (240) converts it into a newsletter format that includes three components: a headline, a summary, and a personalized recommendation. The content generation unit (240) converts it into a learning material format that includes body text (adjustable from A1 to C2), questions, and answers / explanations. The content generation unit (240) converts it into a marketing copy format that can automatically generate main copy, sub-copy, CTA phrase composition, image captions, or titles. The content generation unit (240) converts it into a financial report format that includes a market summary, personalized portfolio analysis, and risk points. The content generation unit (240) sends the personalized content to multiple channels, such as email, apps, messengers, and platform APIs.

[0088] The feedback loop provider (250) collects user response data and reuses it for LLM prompt optimization. The feedback loop provider (250) performs feedback and self-learning.

[0089] The feedback loop providing unit (250) optimizes the prompt based on user response data. The feedback loop providing unit (250) automatically reflects feedback (view rate, correct answer rate, click rate) in the LLM prompt.

[0090] The feedback loop provider (250) improves the quality of reinforcement learning-based newsletters / content. The feedback loop provider (250) generates a reward signal based on the sending result and then operates a model improvement loop based on the reward signal. The feedback loop provider (250) collects user response data to optimize the operation of the AI ​​agent and LLM.

[0091] The feedback loop providing unit (250) according to the present embodiment optimizes the LLM analysis results based on feedback regarding personalized content.

[0092] The feedback loop providing unit (250) reflects feedback, including at least one of the viewing rate, correct answer rate, and click rate for personalized content, into the LLM prompt included in the LLM analysis result.

[0093] The feedback loop providing unit (250) generates a reward signal based on the result of sending personalized content and operates a model improvement loop based on the reward signal.

[0094] The feedback loop provider (250) performs a pre-configured procedure to reflect feedback, such as view rate, correct answer rate, and click rate, into the LLM prompt. The feedback loop provider (250) collects user response data. The feedback loop provider (250) calculates feedback analysis and quality indicators. The feedback loop provider (250) extracts improvement elements (Feedback Tokens). The feedback loop provider (250) automatically reconstructs the LLM prompt, including length / tone adjustment, difficulty adjustment, CTA / keyword optimization, and priority reflection of topics of interest. The feedback loop provider (250) transmits the updated prompt to the LLM analysis unit to improve the generation quality. The feedback loop provider (250) continuously improves the quality of customized content by adjusting the LLM prompt based on feedback.

[0095] The feedback loop providing unit (250) collects feedback data.

[0096] The feedback loop provider (250) collects user responses to personalized content. The feedback loop provider (250) collects the view rate, accuracy rate, and click-through rate (CTR) as feedback values. The feedback loop provider (250) collects the view rate, which includes whether or not the newsletter, report, etc., are viewed and the ratio thereof. The feedback loop provider (250) collects the accuracy rate, which includes the user's accuracy rate in learning materials, quizzes, tests, etc. The feedback loop provider (250) collects the click-through rate, which includes whether or not marketing copy and recommended content are clicked. The feedback loop provider (250) uses the collected feedback values ​​as quantitative indicators of user behavior patterns.

[0097] The feedback loop providing unit (250) calculates feedback interpretation and quality indicators.

[0098] The feedback loop provider (250) analyzes the collected feedback figures to calculate a quality score that evaluates whether the content is suitable for the user. If the viewing rate is low, the feedback loop provider (250) calculates the quality score as the content being too long or too difficult. If the correct answer rate is too high, the feedback loop provider (250) calculates the quality score as the difficulty being too easy. If the click-through rate is low, the feedback loop provider (250) calculates the quality score as the copy text or CTA quality being insufficient. If the click-through rate is high, the feedback loop provider (250) calculates the quality score as the specific phrase, tone, or keyword being effective. The feedback loop provider (250) can configure the quality score as a single value or as a score per item. For example, if the viewing rate is 42%, the click-through rate is 3.8%, and the correct answer rate is 58%, the feedback loop provider (250) calculates the model quality score as 0.61.

[0099] The feedback loop provider (250) extracts prompt improvement elements (Feedback Tokens). The feedback loop provider (250) generates “prompt improvement elements (Feedback Tokens)” to be reflected in the LLM prompt based on the results of the quality indicator analysis.

[0100] The feedback loop providing unit (250) sets the improvement elements in the direction of tone / length adjustment, difficulty adjustment, and marketing copy optimization.

[0101] The feedback loop provider (250) adjusts the tone / length. When the viewing rate drops, the feedback loop provider (250) reduces the length of the text and strengthens the summary level. Conversely, when the viewing rate rises, the feedback loop provider (250) maintains the analytical content.

[0102] The feedback loop providing unit (250) adjusts the difficulty level. If the accuracy rate is lower than the threshold, the feedback loop providing unit (250) lowers the difficulty level (A2 → A1.5). If the accuracy rate is higher than the threshold, the feedback loop providing unit (250) raises the difficulty level (B1 → B2).

[0103] The feedback loop providing unit (250) performs marketing copy optimization. When the click-through rate decreases, the feedback loop providing unit (250) changes the CTA position and text emphasis. When the click-through rate increases, the feedback loop providing unit (250) strengthens the same tone and keywords.

[0104] The feedback loop provider (250) reinforces user interests. If there is a good response to a specific topic, the feedback loop provider (250) prioritizes exposure of that topic. The feedback loop provider (250) excludes topics with no response from the priority of LLM recommendation content. The feedback loop provider (250) converts into LLM prompt sentences based on improvement factors.

[0105] The feedback loop provider (250) automatically reconstructs the LLM prompt. The feedback loop provider (250) automatically reconstructs the previous LLM prompt based on the extracted improvement elements to generate a new LLM input prompt.

[0106] The feedback loop provider (250) updates the 'prompt including user level: B1, category: Tech, style: concise, elements: 3 summaries of the latest tech news' to the 'prompt including user level: B1, category: Tech, style: concise and shorter, elements: 2 summaries of the latest tech news, reinforced 'AI trend' keywords with high click-through rate, length limit: 200 characters or less, low viewing rate of past content → composition of the introduction with an interest-inducing sentence'. If the viewing rate decreases, the feedback loop provider (250) shortens the sentence length and adds an interest-inducing sentence. If the click-through rate related to a specific topic increases, the feedback loop provider (250) expands the proportion of that topic.

[0107] The feedback loop providing unit (250) improves the learning material prompt. When the feedback loop providing unit (250) checks the feedback and finds that the correct answer rate is 30% (too difficult), it improves the prompt to ‘Difficulty: Lower to A2, Sentence length: Average 10 words or less, Vocabulary: Rewrite with words familiar to beginner learners, Question type: 2 multiple choice questions + 1 short fill-in-the-blank question’.

[0108] The feedback loop provider (250) improves the marketing copy prompt. When the feedback loop provider (250) checks the feedback and finds that the CTR is 0.2% (very low), it improves the copy style to be more aggressive and clearer, the benefit-oriented CTA to add a button-type phrase (“Check Now”, “Limited Offer”), the message length to be shortened to 2 sentences, and the product USP to emphasize the keywords “eco-friendly / energy-saving”.

[0109] The feedback loop provider (250) provides an LLM regeneration and iterative improvement loop using the improved prompt. The feedback loop provider (250) ensures that the finally generated prompt is directly applied to the next content generation cycle of the LLM analysis unit (220). The feedback loop provider (250) forms a continuous cyclic structure (closed-loop improvement) of response, analysis, prompt adjustment, regeneration, and re-evaluation.

[0110] FIG. 3 is a diagram illustrating a method for generating real-time customized content according to the present embodiment.

[0111] The customized content creation device (130) collects external data (S310).

[0112] In step S310, the customized content generation device (130) collects the external data by performing real-time crawling on external news, educational materials, product information, and media content.

[0113] The customized content generation device (130) generates an LLM analysis result that performs LLM analysis on external data (S320).

[0114] In step S320, the custom content generator (130) performs an LLM analysis that performs at least one of summarizing, translating, level-based difficulty adjustment, and copywriting of external data.

[0115] The customized content generation device (130) generates a user profile result by performing a user-specific profile based on the LLM analysis result (S330).

[0116] In step S330, the customized content generator (130) performs a user-specific profile including at least one of a learning level, purchase history, and content consumption pattern based on the LLM analysis results.

[0117] The customized content generation device (130) generates and sends personalized content based on LLM analysis results and user profile results (S340).

[0118] In step S340, the customized content generating device (130) generates personalized content including at least one of a newsletter, learning materials, marketing copy, and financial reports based on LLM analysis results and user profile results.

[0119] The customized content generation device (130) transmits personalized content to a terminal (110) using email, an application, a platform API, and a messenger-based multi-channel.

[0120] The customized content generation device (130) optimizes the LLM analysis results based on feedback on the personalized content (S350).

[0121] In step S350, the customized content generating device (130) reflects feedback including at least one of the view rate, correct answer rate, and click rate for the personalized content into the LLM prompt included in the LLM analysis result.

[0122] The customized content generation device (130) generates a reward signal based on the result of sending personalized content and operates a model improvement loop based on the reward signal.

[0123] Although steps S310 to S350 are described as being executed sequentially in FIG. 3, they are not necessarily limited thereto. In other words, since it is possible to modify and execute the steps described in FIG. 3 or to execute one or more steps in parallel, FIG. 3 is not limited to a chronological order.

[0124] As described above, the real-time customized content generation method according to the present embodiment described in FIG. 3 can be implemented as a program and recorded on a computer-readable recording medium. A computer-readable recording medium on which a program for implementing the real-time customized content generation method according to the present embodiment is recorded includes all types of recording devices in which data that can be read by a computer system is stored.

[0125] The above description is merely an illustrative explanation of the technical concept of the present embodiment, and a person skilled in the art to which the present embodiment belongs would be able to make various modifications and variations within the scope of the essential characteristics of the present embodiment. Accordingly, the present embodiments are intended to explain, not limit, the technical concept of the present embodiment, and the scope of the technical concept of the present embodiment is not limited by these embodiments. The scope of protection of the present embodiment shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present embodiment. Explanation of the symbols

[0126] 110: Terminal 120: Network 130: Customized content generator 210: Data Collection Unit 220: LLM Analysis Department 230: User Profiling Section 240: Content Creation Section 250: Feedback loop provider

Claims

Claim 1 A data collection unit for collecting external data; an LLM analysis unit for generating LLM analysis results by performing LLM analysis on the external data; a user profiling unit for generating user profile results by performing user-specific profiling based on the LLM analysis results; and a content creation unit for generating and sending personalized content based on the LLM analysis results and the user profile results. and includes a feedback loop providing unit that optimizes the LLM analysis results based on feedback regarding the personalized content described above. The LLM analysis unit performs text refinement on the external data by removing HTML tags, removing duplicate content, and automatically correcting missing or erroneous sentences; performs semantic unit segmentation on the external data by reclassifying news, product information, and educational materials by paragraph or topic; attaches metadata to the external data by adding one or more of category, language, extraction time, and data reliability score; when generating different prompt templates according to the analysis type to configure the LLM analysis prompt, it configures a summary prompt template to automatically select prompts based on key sentence extraction, structured summarization, and highlight generation; configures a translation prompt template to include country-specific cultural context, tone, and domain terminology in the translation prompt template; configures a difficulty adjustment prompt to automatically adjust sentence length, vocabulary difficulty, and grammatical structure within the difficulty adjustment prompt; configures a copywriting prompt to reflect product characteristics and user profiles; and when inputting the generated prompts and preprocessed text into the LLM engine to perform summarization, translation, level-based difficulty adjustment, and copywriting, key sentence extraction, Summarizes preprocessed data by performing semantic compression and user-specific length adjustment, reflects cultural context and domain characteristics when performing multilingual translation, and automatically generates text suitable for learner levels in the education field by readjusting sentence length, vocabulary, and grammatical complexity according to CEFR levels,For post-processing, verification of sentence naturalness, application of prohibited words or brand tone guides, removal of machine translation traces, and removal of redundant summaries or similar expressions are performed; the user profiling unit integrates LLM analysis results, user behavior data, and internal system data to generate profile input vectors for each user; calculates the learning level of individual users by combining text difficulty and user learning response data from the LLM analysis results; when calculating the difficulty-based learning level, if the correct answer rate increases when consuming content of a specific difficulty level, it is judged as a candidate for a higher level; if the dwell time is longer than a threshold and the error rate increases, it is judged as a candidate for a lower level; corrects the level by analyzing changes in learning patterns during repeated learning of the same topic; determines the user's product preference by combining purchase data and LLM analysis; analyzes purchase history including price range trends, category preferences, and brand loyalty by combining purchase data and LLM analysis; extracts interest factors based on LLM keywords extracted from product descriptions; generates profiling results including customers who prefer premium electronics, customers sensitive to eco-friendly products, and customers with strong repeat purchase patterns; and when extracting content consumption patterns (features), viewing time, click-through rate (CTR), scroll depth, and When analyzing consumption patterns based on dwell time, reading breakpoints (mid-step exit patterns), preferred content by time of day, and consumption trends by device (mobile / PC), and when performing User Embedding and clustering, integrating learning level, purchase history, and consumption patterns into a single User Embedding (User Vector), generating a profile vector consisting of a learning level index, category preference, purchase pattern index, consumption pattern index, time-of-day response pattern, and LLM keyword matching score, classifying the profile vector into profile groups using clustering algorithms (K-means, DBSCAN, etc.) or rule-based clustering, and generating personalization rules based on LLM analysis results,A customized content generation device characterized by generating a personalization rule to be used based on a derived profile, and, when the content generation unit determines the type of content required for a user based on the user profile, prioritizing the generation of learning materials if the user is a learner, determining marketing copy or product recommendation newsletters as the required content type if the user is an e-commerce user, determining financial reports or market data summaries as the required content type if the user is an investment-interested user, and determining news / trend-based newsletters as the required content type if the user is a media consumer. Claim 2 A customized content generation device according to claim 1, wherein the data collection unit collects external data by performing real-time crawling on external news, educational materials, product information, and media content. Claim 3 A customized content generation device according to claim 1, wherein the LLM analysis unit performs the LLM analysis, which includes at least one of summarizing, translating, level-based difficulty adjustment, and copywriting of the external data. Claim 4 A customized content generation device according to claim 1, wherein the user profiling unit performs the user-specific profile including at least one of a learning level, purchase history, and content consumption pattern based on the LLM analysis results. Claim 5 A customized content generation device according to claim 1, wherein the content generation unit generates personalized content including at least one of a newsletter, learning materials, marketing copy, and financial reports based on the LLM analysis results and the user profile results. Claim 6 A customized content generation device according to claim 1, wherein the content generation unit transmits the personalized content to a terminal using email, an application, a platform API, or a messenger-based multi-channel. Claim 7 A customized content generation device according to claim 1, wherein the feedback loop providing unit reflects the feedback, which includes at least one of a viewing rate, a correct answer rate, and a click rate for the personalized content, into an LLM prompt included in the LLM analysis result. Claim 8 A customized content generation device according to claim 1, wherein the feedback loop providing unit generates a reward signal based on the sending result of the personalized content and operates a model improvement loop based on the reward signal. Claim 9 delete

Images