RetainIQ: Churn Prevention Copilot on AMD

RetainIQ is a churn-prevention copilot for online stores. It tells a retention manager which customers are about to stop buying, why, and what to do about each of them within a fixed budget. The problem. E-commerce stores lose revenue quietly: customers don't cancel, they just stop ordering. The warning signs, a late delivery, an angry support ticket, a two-star review in Spanish, are scattered across systems and languages, and nobody reads them all. Retention teams end up sending blanket discounts and wasting budget on customers who would have stayed anyway. What RetainIQ does. Reads every message on AMD. An LLM, DeepSeek-V4-Flash on AMD Token Factory, reads about 6,100 support tickets and reviews in English, Spanish and German, 50 per request and within the rate limit, extracting sentiment, complaint type and explicit threats to leave. 99% of signals come from the LLM. Predicts churn and explains it. A classical logistic-regression model combines order history with those signals. Text from AMD lifts ROC-AUC from 0.734 (orders only) to 0.808. Every score is explained per customer. Plans the campaign. An optimizer picks the best action per customer: a personal call, a discount or nothing. In the demo store it finds $27,218 at risk across 73 high-risk customers. A $500 budget reaches 57 customers and saves an expected $2,380 in revenue, a 99% ROI. Answers questions in plain English. A Copilot, running on AMD, calls these tools and shows a step-by-step trace. Its numbers always come from the model and optimizer, never from guesses. It suggests follow-ups and drafts win-back messages in each customer's language. Built for real teams: accounts with admin, analyst and demo roles saved conversations per-user daily AI limits that protect the AMD quota PostgreSQL guided tours with narration mobile layouts The same code runs on an AMD Instinct MI300X with ROCm and vLLM by changing one setting. Live demo: "Try the demo": https://retainiq-amd-churn-copilot.vercel.app