Real-Time Market Insights with GenAI: How I Built a Market Analysis Agent Using Claude and Bedrock

We’ve all read financial headlines like “Gold prices surge amid trade tensions” or “Oil dips after OPEC+ announcement.” But what if we could turn this flood of market news into something structured and actionable — in real time?

‍

That was the goal behind our latest project at Colibri Digital: to build a Market Analysis Agent that extracts sentiment, emotion, and confidence from financial news articles. It started as an internal tool for commodities, but quickly evolved into a scalable platform that could serve any industry where real-time opinion matters — from finance to pharmaceuticals to retail.

‍

Here’s how it works, why it matters, and how we see this evolving into a commercial offering.

‍

🚀 The Problem: Too Much Information, Not Enough Insight

‍

Every day, thousands of financial articles are published about commodities, stocks, policies, and trends. But stakeholders — whether they’re commodity traders, strategy consultants, or brand managers — don’t have time to read them all.

‍

Manually reading articles to extract “market mood” isn’t just slow. It’s subjective, inconsistent, and impossible to scale.

‍

So we asked ourselves: What if a GenAI agent could read hundreds of articles and deliver structured insights and forecasts in seconds?

‍

‍

🧠 The Solution: GenAI-Powered Sentiment Summarisation

‍

The agent works like this:

‍

1. Input: A user uploads a CSV/XLSX file or retrieves real-time news via NewsAPI filtered by commodity, source, and date range.
   
   
2. Processing: A Django backend parses the content and sends it to Claude 3.7 Sonnet via AWS Bedrock, using a custom prompt to extract sentiment, confidence, emotion, and a summary per commodity. We chose Claude via Bedrock because of its:

‍

🔹 Reliable JSON output for structured parsing

🔹 Lower hallucination rate than other models

🔹 Seamless integration with AWS services

🔹 Flexible prompt control with Claude’s strong summarisation capabilities.

‍

3. Output: The frontend (built in Streamlit and Next.js) displays a clean table and visualisations — with sentiment trends, forecasts, and per-article insights.

‍

What makes it powerful is that the LLM returns structured JSON, which is parsed into a user-friendly dashboard.

‍

‍

‍

‍

‍

💻 Tech Stack (Built to Scale)

‍

This started as an internal tool but is designed to scale across use cases and clients. Here’s what powers it:

• Frontend:
  
  🔹 Streamlit (for internal testing and quick demos)
  🔹 Next.js (production-ready UI in progress)
  
  
• Backend:
  
  🔹 Django REST API
  🔹 Claude 3.7 Sonnet via AWS Bedrock
  
  
• Monorepo:
  
  🔹 Hummingbird AI — our unified repo for AI demos and internal tooling
  
  
• Future Steps:
  
  🔹 Pinecone Vector DB + Bedrock Knowledge Bases for long-term news memory and retrieval
  🔹 Guardrails for output validation and safety.

‍

‍

📊 Why It Generalises Across Industries

‍

AAlthough the initial use case focused on commodities like Gold, Oil, and Gas, we built the system to generalise across domains. Why?

‍

Because the key components — article ingestion, LLM-based summarisation, structured output, and trend tracking — work regardless of subject matter.

‍

By adjusting:

‍

• The data source (e.g. Reddit for retail, clinical trial feeds for pharma, etc.)
  
  
• The prompt context
  
  
• The filters and dashboard layers

‍

…we can rapidly deploy this architecture across verticals.

‍

🔄 Use Cases Across Industries:

‍

• Finance — Crypto, equities, commodities sentiment
  
  
• Retail — Consumer reactions to new products or brand launches
  
  
• Pharma — Public perception on drug rollouts and trials
  
  
• Healthcare — Reactions to NHS policies, insurance reforms
  
  
• ESG & Energy — Sentiment around renewables, climate action, green finance

‍

In short: anywhere public opinion moves markets, this tool is useful.

‍

‍

🛠️ Architecture in Plain English

‍

The flow is simple and modular:

‍

• Streamlit or Next.js frontend handles file upload and NewsAPI retrieval
  
  
• Django backend routes the article content to Claude
  
  
• Claude returns structured sentiment JSON
  
  
• The frontend displays the result in a live dashboard.

‍

We’re also exploring embedding and storing articles in Pinecone via Bedrock’s new Knowledge Base capability — a step toward persistent memory and semantic search.

‍

‍

🧠 Final Thoughts: Why This Matters for Your Business

‍

Whether you’re tracking public reaction to a new policy, understanding how analysts feel about a stock, or gauging sentiment around a brand — Gen AI can unlock speed, scale, and objectivity.

‍

This project shows how companies can go from raw text to structured insights and forecasts in seconds. No manual tagging. No noise. Just clear, confident sentiment at your fingertips.