Intelligent Inventory Forecasting for Hun**** Industries, AustraliaCase Study

Instinct worked for the fastest-moving lines but broke down across the long tail. The business was carrying significant capital in stock that had not moved in over a year, while periodically running out of fast movers and paying for emergency air freight to recover. With supplier lead times measured in months, a poor reorder decision compounded quietly before anyone could see it, and correcting it took just as long.

The underlying data was the harder problem. The ERP held years of transactions, but stock adjustments, returns and inter-site transfers had been recorded inconsistently, so raw consumption figures could not be trusted for forecasting. Seasonality was real but irregular, tracking construction activity rather than the calendar, which defeated simple seasonal models.

There was also legitimate organisational scepticism. The purchasing manager had decades of experience and no appetite for a black box telling him he was wrong. Explainability was therefore a functional requirement: any recommendation had to show its reasoning and history clearly enough for an expert to evaluate it, or it would not be used.

PapaSiddhi began with a data reconciliation exercise before any modelling, classifying historical transaction types in SQL Server and separating genuine consumption from adjustments, returns and transfers. That groundwork took longer than the modelling itself and is what separates a useful forecast from a confident wrong one. Forecasting then runs as a Python service producing per-SKU demand projections with explicit confidence intervals.

Rather than forcing every product through one algorithm, steady movers, intermittent-demand lines and genuinely erratic SKUs are modelled separately. Recommended reorder points factor in both average demand and the variability of supplier lead times. The service reads from and writes recommendations back to the ERP through its REST API, so purchasing continues in the system the team already uses. Forecasts regenerate nightly and every run is versioned, allowing past recommendations to be compared with actual outcomes.

Two interfaces serve different audiences. Power BI, published through the Power BI Service, provides management reporting on stock health and forecast performance, while a React dashboard supports the purchasing team's daily work, showing the history and reasoning behind every recommendation. A manual override path records who changed a reorder point and why, keeping the purchasing manager in control while creating an audit trail the business did not previously have.