FinSight AI
Multi-agent financial-intelligence platform built on a streaming medallion lakehouse.
Problem
Financial analysis spans market data, filings, news, and risk signals that arrive continuously and in different shapes. A single model cannot reason across all of them reliably, and analysts need both fresh features and explainable, tool-using reasoning.
Solution
Six specialized ReAct agents coordinated with LangGraph reason over a streaming lakehouse. Raw events are refined through a medallion architecture into a feature store, then served through an API and an interactive dashboard for analysts.
Architecture
Kafka ingests events serialized with Avro. Spark processes them through bronze, silver, and gold medallion layers into Delta Lake. Feast materializes features into Redis for low-latency online serving. Six ReAct agents orchestrated by LangGraph perform retrieval and tool use, exposed through FastAPI with a Streamlit front end.