financial-analysis-agent

The Financial Analysis Agent is an FP&A copilot that answers plain-English questions about company financials by pulling data from SEC filings and market sources, running the analysis in Python, and explaining the results with sources attached. It is a Python application that users run on their own computer using their own Anthropic API key, with no hosted version available. The system is designed to avoid the common problem of LLMs generating plausible but ungrounded financial figures: the model decides what to look up and how to explain it, but every number shown is the output of a deterministic Python computation over real SEC filings, with its source attached. The LLM never does arithmetic. The agent can answer questions about public companies using SEC EDGAR data, or users can upload their own small business's numbers as a CSV and ask the same kinds of questions about their own data. The project includes an evaluation harness that tests grounding, with results showing a pooled strict grounding rate of 93.9% across 21 eval runs. The system is built with a multi-layer architecture: an agent layer that decides which tools to call, a data layer for SEC EDGAR and market data, an analysis layer for deterministic computations, and guardrails for source attribution and consistency checks. It is a research and educational project, not financial advice.