Sales Data Synthesizer Development
Explore how Celadon's idea to create a sales data synthesizer helped train an effective AI model. Discover the impact of this breakthrough in our case study.
Project Overview
| Industry | eCommerce |
|---|---|
| Duration | 3 months |
| Services |
|
Technology Stack Used
Challenge
- Sparse data: Limited historical transactions created a cold-start problem for most products and users.
- Real-time speed: Recommendations had to appear in under 150 ms at cart time.
- Seasonal swings: Flash sales and holidays altered buying patterns, risking model drift.
- Drop-in integration: New service needed to fit the existing JSON API with no backend refactor.
Solution
- Sales-Data Synthesizer that spun a small seed of real transactions into millions of realistic records
- Collaborative filtering model trained on the synthetic + original data blend
- Real-time workflow:
- Shopper adds item to cart
- Model scores correlations across synthetic cohorts
- System returns highly targeted cross-sell suggestions within milliseconds
Impact for the client
• Correlation hunting now includes “hidden” signals absent from the thin live dataset
• Recommendation click-through and add-to-cart rates jumped, boosting average order value
• Marketplace gains an extensible AI asset without waiting years to accumulate raw data