Transforming the Retail Industry with Transformers

Abstraction: BERT transformers automating retail product attribute classification in production

Key points:

• Georgian partnered with SPINS (retail product data company) to automate attribute prediction (brand, unit size, etc.) from noisy retailer product descriptions across 86,000+ brands.
• Rule-based model achieved 77% accuracy on brand identification but lacked scalability and confidence scores; BERT fine-tuned classifier achieved only 55% accuracy on brand but was productionized for its scalability and calibrated probabilities.
• Unit size prediction was framed as a sequence-to-sequence translation problem using HuggingFace Seq2Seq, achieving ~84% accuracy.
• Serving stack: Docker images served via REST API on Google Cloud Run (8GB RAM, 4 CPU); horizontal scaling to millions of predictions per hour; model assets stored in Google Cloud Storage as an internal model registry.
• Retraining pipeline: triggered every four weeks on new data, managed by Apache Airflow; shadow testing before promoting models to production; performance monitored via a dashboard.

Connections: Bert · Spins · Georgian · Transformers · Natural Language Processing · Mlops

Source: https://medium.com/georgian-impact-blog/transforming-the-retail-industry-with-transformers-f429f792474a