juanjuandog/finsight-ai

FinSight AI is an open-source tool that turns company filings, financial reports, and market data into AI-generated equity research reports. Instead of giving you a plain chat answer, it produces structured briefs with ratings, confidence levels, positive points, and risk points, all backed by citations to the original source documents. It also includes a dashboard where you can search a company, view price trends, and see the AI's investment thesis in one place. The project is designed to show how to build the infrastructure around an AI agent, not just how to call a language model. When you ask it to analyze a company, it runs a multi-step workflow: ingesting documents, calculating financial metrics, indexing text for search, building a knowledge graph of company events, and finally generating a versioned report. Each piece of evidence in the report links back to a specific chunk of a filing, so you can trace every claim. It uses a database that supports similarity search to find relevant document passages, and it caches reports based on the underlying data snapshot, so nothing gets recomputed unnecessarily. Who would use this? A startup founder building a financial analytics product, an engineering manager interviewing for a role that involves AI infrastructure, or a product manager exploring what a dependable AI research pipeline looks like. It's particularly aimed at people who have seen basic AI demos and want to understand what it takes to make one production-ready, with retries, failure recovery, duplicate-task prevention, and quality evaluation built in. What's notable is the focus on trust and resilience. The system tracks whether AI answers might be hallucinating by checking evidence coverage and conclusion consistency. It handles long-running tasks that might fail midway, with automatic retries and dead-letter queues for stuck work. You can run the full stack locally with Docker using sample data, and it works even without a local language model installed, falling back to deterministic rule-based analysis. The README doesn't detail production deployment beyond local setup, but the architecture is clearly designed to be resilient under real-world conditions.