Boosting Steel Manufacturing Profitability with AI Forecasting

Quick Summary

Challenge

High material costs (85% of total) and frequent machine downtime hurt margins due to unreliable forecasts and macroeconomic volatility.

Solution

We combined historical SAP order data, macroeconomic pricing signals, and real-time IoT health metrics using RNN/LSTM and Attention models to forecast demand, predict pricing, and minimize machine downtime.

Result

75% improvement in forecasting accuracy

Tech Stack

AI & ML: PyTorch RNN/LSTM Attention architectures | Data & ETL: SAP ERP SQL Python Web scraping (industry reports and news) | IoT & Infra: Edge sensors MQTT Kubernetes-hosted inference services

The Challenge

A leading cold-rolled steel manufacturer faced raw material costs accounting for ~85% of total production spend. The input materials varied in grade and size, making forecasts difficult and margins thin.

A Day in the Life: Before Our Solution

Every morning, the procurement team logged into SAP to find that raw steel billet prices had shifted again. With material costs making up nearly 85% of the final product cost, even small fluctuations had major financial consequences.

On the shop floor, engineers had to work with input materials that varied in grade and width, throwing production plans into disarray. Unplanned stoppages were common, leading to hours of downtime each month.

Forecasting teams spent countless hours trying to align orders and maintenance schedules across disconnected systems. Despite their best efforts, they kept missing key signals in demand and supply.

Without a smarter forecasting system, the business was stuck in a high-risk cycle — overstocking expensive inputs, under-producing during demand spikes, and losing millions each year to inefficiencies they could see but couldn't prevent.

Solution

To address these challenges, we designed a system that blends predictive modeling with on-the-ground reality.

1. Core Innovation

We integrated three disparate data streams — historical order data, macroeconomic inputs, and live IoT sensor feeds — into a single, intelligent forecasting engine. Using RNNs with attention mechanisms, the models could capture temporal dependencies, seasonality, and event-based anomalies.

2. Key Features

Automated ingestion from SAP, industry publications, and sensor networks
Multi-output models predicting order volume, raw material prices, and machine health
An optimization layer that recommends ideal procurement volumes based on forecasted needs and risk profiles

3. Workflow Integration

The system connects to existing dashboards and planning tools via a lightweight API. Procurement and planning teams can access daily predictions and optimization insights without changing their current workflows.

Outcomes

75% improvement in prediction accuracy for both order volume and machine downtime

Forecasts now run automatically, enabling faster procurement decisions and less dependence on manual analysis