An AI research assistant for capital markets analysts
Analysts spent two to three hours digging through PDFs for a single data point. The answer, with its source, now arrives in under 30 seconds.
To find a cited data point, down from 2–3 hours
From a pilot group of 4 analysts to the broader research team, by word of mouth
- Client
- A capital markets firm
- Users
- Research analysts
- Sources
- Thousands of annual reports (up to 200 pages each) and earnings call transcripts
- Role
- End-to-end delivery
- Service
- Custom AI Agent Development
Biruni Applied AI built an AI Research Assistant for a capital markets firm to help analysts instantly query thousands of annual reports (up to 200 pages each) and earnings call transcripts. The engagement covered the full solution, from data to deployment.
How it was built (the architecture)
Instead of asking the AI to memorize all the data, the system was built like a digital librarian:
The Library (Database): Millions of pages of financial data were organized into a searchable “index” where the AI could look up specific facts, like debt ratios or CEO comments, in seconds.
The Researcher (Search): When an analyst asks a question, the system finds the exact relevant paragraphs first, before the AI ever sees the query. This is the RAG (Retrieval-Augmented Generation) layer.
The Writer (GenAI): The AI then reads only those specific paragraphs to write a concise, cited answer. This prevents the AI from “making things up” because it must use the provided text.
Steps to production-grade (reliability)
To make sure the tool performed at a professional grade:
The Fact-Check: The AI was required to provide footnotes for every claim. If it couldn’t point to the source page, it wasn’t allowed to answer.
Security: The data stayed within the company’s private “vault” so sensitive market insights never leaked to the public internet.
Stress Testing: The system was tested against “trick questions” from senior traders to make sure it wouldn’t get confused by complex financial jargon.
The result
Analysts who used to spend 2–3 hours digging through PDFs for a single data point can now get that answer in under 30 seconds. Adoption expanded from a pilot group of 4 analysts to the broader research team within 6 weeks of launch, driven entirely by word of mouth.
- —Find the source first, then let the AI write. It only sees the paragraphs it needs.
- —No source, no answer. Every claim carries a footnote to the page it came from.
- —Test with your hardest users before launch, not after.
Client name and identifying details are withheld to protect confidentiality. Outcomes are as measured.