Financial Research Agent

Disclaimer

This write-up is for research and education only. It is not investment advice and is not a recommendation to buy or sell any security.

Problem Statement

We were reached out by a client who wanted to build a financial research agent that would help them in their stock research. The principles were relatively simple :

• The agent should be able to gather financial data from the web and tools (fiscal.ai, screener)
• The agent should be able to normalize the financial data
• The agent should be able to create simple DCF models independently
• The agent should be able to generate a report of the stocks with a simple UI (think of a pdf report)

The core gap is not data access. It is a repeatable workflow that moves from prompt to ranked basket to report artifact without losing rigor.

System Design

1. User submits a stock-basket research prompt.
2. The agent fans out retrieval using Exa and parallel calls.
3. Evidence is normalized across ROCE, revenue, P/E, PAT, debt, and margin notes.
4. A second pass builds a quality mental model per business.
5. A sentiment layer adds a 6-12 month directional view.
6. Tiering logic ranks names into clear buckets.
7. Final output is written to TeX and compiled to PDF.

The main design choice is separation of concerns: research, normalization, ranking, and report generation are treated as separate stages.

Lets see how the UX for analyst looks like

User prompt is a simple text input that asks for a stock-basket research prompt.

The agent then fans out retrieval using Exa and parallel calls and uses tools like seaborn, fiscal.ai apis and morningstar.

Please read a sample report generated by the agent here :
AI Boom Stocks Report 2026