PulseCommerce AI

Small retailers lose money in two ways: products run out and sales are lost, or too much stock sits unsold and cash is locked up. Most owners decide by gut feeling. PulseCommerce AI is an agent that tells a merchant what to reorder and what to discount, explains why, and waits for a person to approve before anything happens. How it works: a gradient boosting model forecasts weekly demand per product from sales history, seasonality, price and promotions. We do not use a language model for forecasting, so every number can be measured. Simple rules turn the forecast, stock level and supplier lead time into proposed actions: reorder with quantity and urgency, or discount with a percentage. Gemma 4, served on AMD GPUs through vLLM, explains each action in plain language and drafts the supplier email. Evolus runs the process through its API, with a knowledge library for supplier rules, MCP to update the inventory system, and mandatory manager approval. In a rolling backtest on weeks the model had not seen, it beat both a seasonal-naive baseline and a 4-week moving average (12.4% error versus 13.5% for the best baseline). This test used synthetic data, so it demonstrates the method, not final accuracy. Next steps: real retail data and a live inventory system. Built with Gemma 4, AMD GPUs, vLLM, ROCm, Evolus, Python, scikit-learn and Docker.