Python, FastAPI, React AI: Anthropic Sonnet 3.5, Langchain, LCEL, Langsmith

4 Months

The client specializes in investment fund analysis. The aim of the software is to dramatically reduce the time investment analysts spend reading fund descriptions, presentations and financial data. The solution allows for the creation of condensed and direct analysis of the funding firm, track record of previous funds, etc.

The client needed software for financial and investment analysts that could handle documents in different formats and create customized financial documents, tables, reports, and diagrams. They wanted the software to track changes with undo/redo options and include a chat feature to answer questions and generate document sections based on natural language requests.

The software built for financial and investment analysts intakes a set of documents in various formats and generates customized financial documents, tables, reports, and diagrams.

The program supports history of changes (with undo/redo functionality), and a native chat interface, answering user questions about the financial data and generating document sections upon natural language request. The application uses multiple Anthropic agents, complex workflows and multi-step prompting with Langchain and LCEL.

The main challenge of the project was to analyze multi-page (over 100) documents and make the model produce consistent results over multiple prompts, without using Retrieval-Augmented Generation (RAG). We utilized a huge 200k token context of Sonnet 3.5 and achieved good results in context retention, fact extraction and generation accuracy.