Data entry ops · Production

Kantar Worldpanel uses Databricks and GPT-4 to generate 120,000 training pairs at 94% accuracy

The problem

Kantar Worldpanel's legacy systems were inflexible, resource-intensive, and required an outdated programming skillset, blocking their ability to experiment with modern AI/ML models and scale their data insights business.

Workflow diagram · grounded in source
1
Receipt-to-barcode matching need
Trigger
Pairing receipt line descriptions and product barcode names is the upstream use case that starts the workflow.
▾ source quote
“Pairing receipt line descriptions and product barcode names is just one of the use cases in which we're trying to leverage GenAI”
2
Multi-LLM experimentation
Ai action
Llama, Mistral, GPT-4, and GPT-3.5 are all evaluated within the Databricks Platform; GPT-4 provides the best answers at 94% accuracy.
▾ source quote
“We've experimented with Llama, Mistral, GPT-4 and GPT-3.5, all within the Databricks Platform. Ultimately, GPT-4 provided better answers, with an accuracy of 94%”
3
Automated training data generation
Ai action
GPT-4 automatically generates a training dataset of about 120,000 pairs of receipt descriptions and barcode names at 94% accuracy in a couple of hours.
▾ source quote
“automatically generated a training dataset of about 120,000 pairs of receipt descriptions and barcode names with an accuracy of 94% in just a couple of hours”
4
Smaller model fine-tuning
Ai action
A smaller, more cost-effective model is fine-tuned on the generated training data and served in the production pipeline.
▾ source quote
“we can use it to generate training data to fine-tune a smaller model and serve it in our production pipeline. Smaller models are not only more cost-effective but more performant”
Reported outcome

Kantar Worldpanel automatically generated a training dataset of about 120,000 pairs of receipt descriptions and barcode names at 94% accuracy in a couple of hours, freeing manual coding teams and engineering resources for higher-value work.

Reported metrics
model accuracy (GPT-4 and generated training data)94%
Training dataset size generated120,000 pairs
Time to generate training datasetin just a couple of hours
Resource utilizationsignificant resource optimization
Show all 5 reported metrics
model accuracy (GPT-4 and generated training data)94%
training dataset size generated120,000 pairs
time to generate training datasetin just a couple of hours
resource utilizationsignificant resource optimization
smaller model cost and performancemore cost-effective but more performant
Reported stack
DatabricksMLflowMosaic AI Vector SearchUnity CatalogDatabricks MarketplaceLlamaMistralGPT-4GPT-3.5
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://www.databricks.com/customers/kantar-genai
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Kantar Worldpanel automatically generated a training dataset of about 120,000 pairs of receipt descriptions and barcode names at 94% accuracy in a couple of hours, freeing manual coding teams and engineering resources…

What tools did this team use?

Databricks, MLflow, Mosaic AI Vector Search, Unity Catalog, Databricks Marketplace, Llama, Mistral, GPT-4, GPT-3.5.

What results were reported?

model accuracy (GPT-4 and generated training data): 94%; Training dataset size generated: 120,000 pairs; Time to generate training dataset: in just a couple of hours; Resource utilization: significant resource optimization (source-reported, not independently verified).

How is this data entry ops AI workflow structured?

Receipt-to-barcode matching need → Multi-LLM experimentation → Automated training data generation → Smaller model fine-tuning.

Related data entry ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Data entry ops cases →