NoordSight: How We Built an AI Market Intelligence Agent from Scratch

The problem hypothesis, architecture decisions, mistakes, and lessons from building our first AI agent. A technical behind-the-scenes look.

· ChangeMy.AI Team · Build in Public

NoordSight is our AI market intelligence agent. It analyzes European markets, tracks competitors, identifies potential partners and customers, and delivers actionable market reports — automatically.

This is the story of how we built it from scratch.

The Problem Hypothesis

SMEs make critical business decisions based on incomplete information — not because they don't want better data, but because the tools that provide it are priced for enterprises.

Our hypothesis: An AI agent that continuously monitors public data could deliver 80 percent of the intelligence value at 5 percent of the cost. We validated this with 15 conversations with KMU founders.

Architecture Decisions

LLM Selection

We use a multi-model approach. Claude for complex analysis and report generation. GPT-4 for structured data extraction. Smaller models for routine classification and filtering.

Data Pipeline

NoordSight ingests from public company data, news and media, social signals, and structured databases. The pipeline runs continuously, not in batches.

Agent Architecture

NoordSight is an ecosystem of specialized sub-agents:

Each communicates through a shared context layer.

What Worked (and What Didn't)

Worked: Starting with a narrow scope. European tech market intelligence only. Quality was dramatically better for it.

Worked: The multi-agent approach. Specialized agents were more reliable and easier to debug.

Didn't work: Over-engineering the data pipeline before validating what users actually needed.

Didn't work: Assuming users wanted raw intelligence. They wanted answers: Should I enter this market? Who are my competitors? What partnerships should I pursue?

What's Next

Deeper integration with CRM and sales agents, expansion to additional European markets, predictive capabilities, and custom alert configurations.

The Takeaway

Start narrow, go deep. Talk to users before you build. Use the right model for each task. Build modular. Ship early — real-world feedback is worth a thousand assumptions.