.png&w=256&q=75)
🤓 Latest Submissions
LastCall
The problem Most shops lose customers quietly. A customer stops buying, and nobody notices until it is too late. Winning back an old customer is cheaper than finding a new one, but teams do not know who to call or what to say. What LastCall does LastCall looks at real purchase history and finds the customers who are about to leave. It gives each customer a risk score from 0 to 100. Then an AI writes a personal win-back message for each one: why the customer may be leaving, which offer fits, and a short ready-to-send email. How it works A Random Forest model reads 20 signals about each customer (for example, days since the last order compared to their usual rhythm) and gives the risk score. Qwen3.8-27B, running on AMD Token Factory (AMD GPUs), writes the reason, the offer type and the message, in one call per customer. A simple dashboard shows the customers by risk, with the offer for each one. Results on real data 4,317 active customers scored. On customers the model had never seen, the high-risk group (score 70-100) was right: 77.5% of them really stopped buying. We focus the offers on 1,653 customers worth contacting: customers with at least 2 orders and high risk, plus one-order customers who spent more than 500 £. 200 personal offers were written in 23.5 minutes for about 0.03 $ in total. Why it matters Small shops cannot afford a data team. LastCall turns purchase history into a clear list of who to contact and what to say, at a very low cost. Demo note The online demo shows the dashboard with the offers we already generated. Live offer generation needs our private API key, so it only runs locally; we show it in the video.
18 Oct 2026
