Forecasting & segmentation

The problem

Merchants at Zeal were flying blind. They had transaction history but no forward view of how many customers would walk in or how much they would spend, so staffing, stock, and promotions were all guesswork. Meanwhile the signals that revealed merchant health, statistical indicators, RFM behaviour, predictive scores, lived in separate dashboards nobody could reconcile into a decision.

The approach

1. 01

   Built demand forecasting end to end

   Modelled customer visits and spend per merchant, from raw transaction pipelines through feature engineering to a served forecast with a usable horizon and confidence bands, so merchants could plan against a reliable signal instead of intuition.

2. 02

   Fused indicators into one health score

   Combined RFM segmentation with statistical and predictive indicators into a single 0-100 merchant health score, weighting each signal so the number stayed decision-ready rather than another chart to interpret.

3. 03

   Made LLM output production-grade on GCP

   Engineered and hardened prompts so LLM-generated summaries and explanations were consistent, grounded, and safe to ship inside the product on GCP, not just demo-quality.

4. 04

   Used AutoML for the hard tabular cases

   Where tabular forecasting got too complex for a single hand-tuned model, I leaned on AutoML to search the model space, then validated and framed the winners so the forecasts held up in production.