Ecommerce ops · Production

Delivery Hero QC automates product attribute extraction and title standardization with predefined agentic AI

The problem

Delivery Hero QC's manual processes for verifying and enriching product attributes across large, growing catalogs were time-consuming, costly, error-prone, and could not scale across numerous platforms and geographical regions.

Workflow diagram · grounded in source
1
Vendor title and image ingested
trigger
“The first LLM receives the vendor product title and image”
2
Attribute extraction by LLM
ai_action
“Its task is to extract a comprehensive set of 22 predefined attribute types (Brand, Flavor, Volume, etc.)”
3
Standardized title generation
ai_action
“The second LLM receives the extracted attributes, and its task is to generate a new product title that follows the standard Delivery Hero QC format”
4
Confidence scoring gate
validation
“processing the output logits from the LLMs and converting them into probability scores, effectively quantifying the model's confidence in its generated attributes and titles. Outputs falling below a predefined confidence threshold are au…”
5
Human review of flagged outputs
human_review
“ensuring that edge cases or potentially uncertain predictions receive manual verification”
6
Product knowledge base enriched
output
“extracting the nuanced details we need to enrich our product knowledge base”
Reported outcome

Delivery Hero QC is successfully automating product attribute extraction and title standardization using LLM-powered AI agents, delivering improvements in efficiency, accuracy, data quality, and customer satisfaction, with significantly lowered operational costs and latency through knowledge distillation.

Reported metrics
Catalog enrichment efficiencyReducing manual effort and speeding up catalog enrichment
Product information accuracyEnsuring consistent and correct product information
Operational costs and latencysignificantly lowered operational costs and latency
Customer satisfactionmore reliable and easily navigable product discovery experience
Reported stack
LLMsGPT-4oGPT-4o-mini
Source
https://tech.deliveryhero.com/blog/how-delivery-hero-uses-agentic-ai-for-building-a-product-knowledge-base/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Delivery Hero QC is successfully automating product attribute extraction and title standardization using LLM-powered AI agents, delivering improvements in efficiency, accuracy, data quality, and customer satisfaction,…

What tools did this team use?

LLMs, GPT-4o, GPT-4o-mini.

What results were reported?

Catalog enrichment efficiency: Reducing manual effort and speeding up catalog enrichment; Product information accuracy: Ensuring consistent and correct product information; Operational costs and latency: significantly lowered operational costs and latency; Customer satisfaction: more reliable and easily navigable product discovery experience (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Vendor title and image ingested → Attribute extraction by LLM → Standardized title generation → Confidence scoring gate → Human review of flagged outputs → Product knowledge base enriched.

Related ecommerce ops cases