97% Accurate Product Categorization at Scale
How we built an ML pipeline that classifies marketplace products into 5,585 categories with 97% accuracy — replacing a manual process that could not keep pace with inventory volume.
The Challenge
A large online marketplace had a categorization problem that only got worse as they scaled. With 5,585 product categories and an ever-growing product catalogue, manual classification was a full-time job that could never catch up. Miscategorised products meant poor search results, worse SEO, frustrated buyers, and lost sales.
Previous automated attempts had hit a ceiling around 70–75% accuracy — good enough to feel like progress, but not good enough to replace human review. The team was still manually checking thousands of classifications every week.
What We Built
We built a machine learning pipeline that classifies products into the full 5,585-category taxonomy with 97% accuracy — well above the threshold needed to run without manual review on the vast majority of items.
The system combines product title, description, and attribute data as input signals, processes them through a fine-tuned classification model, and returns a category assignment with a confidence score. High-confidence predictions are applied automatically. Low-confidence predictions are routed to a human review queue — keeping accuracy high while eliminating wasted effort on straightforward cases.
How It Works
- Product data ingestion — title, description, and available attributes pulled for each product
- Feature engineering — text signals normalised and structured for the classification model
- ML classification — fine-tuned model predicts the most likely category from 5,585 options
- Confidence scoring — each prediction returned with a confidence score
- Routing logic — high-confidence predictions applied automatically; low-confidence flagged for human review
- Feedback loop — human corrections fed back into the model to improve accuracy over time
Need Help Building AI Automations?
The Results
- 97% classification accuracy across 5,585 categories
- Majority of products categorised automatically without human review
- Human review queue reduced to only genuinely ambiguous cases
- Improved product discoverability and marketplace search relevance
- Model improves continuously as corrections are fed back into training
