Retail · AI in production
Millions of reviews, asked a question in plain English
Topic modelling over millions of customer reviews, with a conversational agent on top so the answer takes seconds rather than an analyst.
- Client
- Brand reputation software provider
- Segment
- Retail, customer experience
- Training sample
- about 1 million reviews
- Deployed on
- Amazon SageMaker Pipelines
- Sentence Transformers
- BERTopic
- MiniBatch K-Means
- c-TF-IDF
- Amazon SageMaker
- Kubernetes
Before
What was true before
The client had collected large volumes of feedback, ratings and sentiment, and its dashboard showed the numbers without what customers were talking about. Sending millions of reviews straight to a language model hit context limits, latency and inconsistent output, and made comparisons across time unreliable.
After
What we did
A topic-modelling pipeline built for their data: sentence embeddings, clustering with stratified sampling, a topic hierarchy, and topic quality measured rather than assumed. A conversational agent sits over it and drills down to the reviews behind each answer.
Build log
How it was built
- 01
Prototyped with BERTopic, then replaced it with a custom topic-modelling pipeline
- 02
Turned reviews into semantic vectors with Sentence Transformers
- 03
Clustered with MiniBatch K-Means and stratified sampling to build a training set of about 1 million reviews
- 04
Built a topic hierarchy and a custom distinctness metric to judge topic quality
- 05
Named topics with c-TF-IDF and joined them to the existing sentiment data
- 06
Built a dashboard and a conversational agent that answers questions and shows the supporting reviews
- 07
Deployed on Amazon SageMaker Pipelines with asynchronous inference beside the existing sentiment pipeline
Measured
Results
As reported in the published case.
- or less to answer a question
- 15 seconds
- to deliver the conversational agent
- 4 months
- reported
- Millions of reviews analysed
Engagements
The services behind it
Next step
Yours would be the next one.
Tell us where it breaks, and you get a written answer on what breaks first and what it costs.