Ecommerce ops · Production

GoDaddy builds a scalable AI product categorization system using Amazon Bedrock batch inference

The problem

GoDaddy's existing product categorization used an out-of-the-box Meta Llama 2 model across six million SKUs, but the generated categories were often incomplete or mislabeled, and running individual LLM calls per product was too costly at scale.

First attempt

The out-of-the-box Meta Llama 2 approach produced categories that were frequently incomplete or mislabeled, and the per-product API call approach was prohibitively expensive for large-scale deployment.

Workflow diagram · grounded in source
1
Product JSONL uploaded to S3
trigger
“A JSONL file containing product data is uploaded to an S3 bucket, triggering the first Lambda function”
2
Lambda spins up Bedrock endpoint
integration
“The Lambda function spins up an Amazon Bedrock batch processing endpoint and passes the S3 file location”
3
Bedrock generates categorized output
ai_action
“It reads the product name data and generates a categorized output, including category, subcategory, season, price range, material, color, product line, gender, and year of first sale”
4
Output written to S3
output
“It writes the output to another S3 location”
5
Lambda monitors and terminates job
integration
“It monitors the batch processing job on Amazon Bedrock. It shuts down the endpoint when processing is complete.”
6
LangChain parses output to CSV
output
“This file is parsed using LangChain's PydanticOutputParser to generate a .csv file”
7
Human evaluation by GoDaddy SMEs
human_review
“The categories generated were evaluated qualitatively by GoDaddy SMEs. The categories were found to be of good quality.”
Reported outcome

The batch inference solution achieves 97% category coverage on both the 5,000 and 100,000 hold-out test sets, processes 5,000 products in 12 minutes (80% faster than the 1-hour requirement), and is 8% more affordable than the Llama2-13b proposal while providing 79% more coverage.

Reported metrics
Product catalog sizesix million products
Solution processing time for 5,000 products12 minutes
GoDaddy maximum processing time requirement5,000 products in 1 hour
Processing latency improvement vs requirement80% faster
Show all 12 reported metrics
product catalog sizesix million products
solution processing time for 5,000 products12 minutes
GoDaddy maximum processing time requirement5,000 products in 1 hour
processing latency improvement vs requirement80% faster
solution cost vs Llama2-13b proposal8% more affordable
category coverage vs Llama2-13b proposal79% more coverage
category coverage on hold-out test sets97%
computation time scaling from 5k to 100k sampleseight times
inference time increase for individual vs batch (100k products)approximately 40 times
format parsing latency reduction via prompt engineeringapproximately 77%
near real-time inference per product2 seconds per product
qualitative category quality assessmentof good quality
Reported stack
Amazon BedrockLangChainPydanticOutputParserOutputFixingParser
Source
https://aws.amazon.com/blogs/machine-learning/how-godaddy-built-a-category-generation-system-at-scale-with-batch-inference-for-amazon-bedrock?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The batch inference solution achieves 97% category coverage on both the 5,000 and 100,000 hold-out test sets, processes 5,000 products in 12 minutes (80% faster than the 1-hour requirement), and is 8% more affordable…

What tools did this team use?

Amazon Bedrock, LangChain, PydanticOutputParser, OutputFixingParser.

What results were reported?

Product catalog size: six million products; Solution processing time for 5,000 products: 12 minutes; GoDaddy maximum processing time requirement: 5,000 products in 1 hour; Processing latency improvement vs requirement: 80% faster (source-reported, not independently verified).

What failed first in this deployment?

The out-of-the-box Meta Llama 2 approach produced categories that were frequently incomplete or mislabeled, and the per-product API call approach was prohibitively expensive for large-scale deployment.

How is this ecommerce ops AI workflow structured?

Product JSONL uploaded to S3 → Lambda spins up Bedrock endpoint → Bedrock generates categorized output → Output written to S3 → Lambda monitors and terminates job → LangChain parses output to CSV → Human evaluation by GoDaddy SMEs.

Related ecommerce ops cases