SITUATION

Domain: Multi-dealer marketplace

Environment: Web + internal tooling Scale: 300+ dealers, 100k+ listings

What Was Unclear

• Was the system answering questions or driving buying intent?
• What constitutes a "good" answer?
• Which constraints matter most: speed, accuracy, recall, cost?

Reframed As

Not an Al chatbot. An inventory discovery engine with a conversational interface.

Problem

Users asked natural language questions.

System returned answers.

But answers did not reliably surface purchasable inventory.

SYSTEM STRUCURE

Intent Layer

• Why: Faster iteration, from aption instead of fine-tuning

Key Decisions

• Used retriev/-augmented generation instead of fine-tuning
  Why: Faster iteration, less upfront cost
• Centralized schema for inventory attributes
  Why: Consistent filtering, easier onboaanding for new dealers
• Separated language handing from business logic
  Why: Simplifies multilingual support, reduces maintenance