The client is a company specializing in data science and reputation management. They have developed an innovative product that integrates academic evolutionary biology and automated data visualization to improve the speed and accuracy of reputation management.

An essential component of the product is a set of metrics and algorithms for calculating the numerical characteristics of the client’s reputation. Some reputation indicators rely on analyzing media content associated with the company. The client aims to provide its clients with monthly reports on reputation changes, detailed explanations, and actionable advice.

However, the media content supplied by a third party was often generalized, making it difficult to analyze and categorize. The generalized content was often irrelevant or lacked the context needed to link the article to the specific company or its reputation.

Before creating the report, we identified several problems related to the quality of the supplied content and its duplication. To address these issues, we:

• Developed Proprietary Algorithms: We stopped using third-party articles and developed our own AI-based summarization and categorization algorithm.
• Correlated Metrics with Article Categories: We used article categorization to align with changes in relevant metrics.
• AI-Generated Reports: We used AI to generate reports and recommendations based on the provided set of metrics and excerpts from relevant articles.
• Vectorization for Duplication Detection: We employed vectorization of article text to identify duplications or reposts before including them in the report.

Our solution utilized the Amazon Bedrock platform with the following foundation models:

• Anthropic Claude: For report generation.
• Amazon Titan: For summarization, categorization, and vectorization.

We used the Langchain.js framework running under Node.js, which decoupled the main logic from specific APIs and simplified work with various aspects of LLMs. Vector storage was implemented within the application’s main DBMS (PostgreSQL), utilizing the pgvector plugin.

• Improved Relevance: Switching to a proprietary algorithm enhanced article relevance and text generation for personalized customer dashboards.
• Comprehensive Reports: Correlating changes in metrics with article categories and using deduplicated texts resulted in understandable and relevant reports.
• Cost-Effectiveness: Leveraging multiple models on Amazon Bedrock with Langchain.js reduced overall solution costs.