Thomas K.

Portfolio

Halcor

A product search system that understands messy, incomplete or bilingual descriptions and finds the right item more accurately.

Role
AI Engineer, Netcompany
Status
In use
Technology (for technical readers)
LangGraph · vector search · GPT-4o-mini · LLM-as-a-judge
+23%improvement in matching F1 score
3steps to find the best match
11bilingual examples used for quality
How it works
Understand the request
Find products with similar meaning
Check sizes and specifications
Compare the most promising results

Select the best match23% improvement in F1 score; live in production

If the search is unclear, try a broader search once

What I built
  1. Improved the product-matching F1 score by 23% compared with the original approach by combining meaning-based search, specification checks and AI review. The system is live.
  2. Designed a backup search process for unclear or unusual requests, helping the system recover when an initial search fails.

Could AI solve a problem for your team?

Get in touch