The workflow before
OptimizeRx's pharma clients wanted price forecasts across therapeutic categories, but existing methods missed by 25 to 30%. The daily data refresh took 6 to 8 hours in batch, business insights trailed by 24 to 48 hours, and onboarding a new client's data took 4 to 6 weeks of integration work.
What we built
We built a price forecasting service that runs Prophet, Amazon Forecast and custom gradient-boosted models side by side, backtests them, and picks or blends the best per series. It produces 30, 60, 90 and 180 day forecasts with confidence ranges and retrains automatically as new data lands.
We moved the data pipeline to Azure: Azure Data Lake Storage with Delta Lake in bronze, silver and gold layers, Azure Databricks for Spark processing and Azure Synapse for the warehouse, with Power BI on top.
Models are tracked in MLflow and trained and served through Amazon SageMaker, so a new model version goes live in days rather than weeks.
Before and after
| Measure | Before | After |
|---|---|---|
| Daily data refresh | 6 to 8 hours | 45 to 60 minutes |
| Price forecast error (MAPE) | 25 to 30% | 12 to 15% |
| New client data onboarding | 4 to 6 weeks | 1 to 2 weeks |
| Model deployment | 2 to 3 weeks | 2 to 3 days |
What's next
Forecasts and clean data in one place set up the next step: putting forecasts and their drivers in front of client teams through a copilot that answers pricing questions directly.
