wtf is churnpilot?

ChurnPilot is an AI powered platform for predicting which customers are likely to cancel a subscription or service, and for automatically taking retention action before they do. The project description is written in Turkish and translates to agentic AI customer churn prediction and automated retention platform. This repository is the early foundation of the project, described as Sprint 1 through Sprint 4 of a larger build, with a working full stack skeleton, a complete database schema, and a routing shell already in place. The backend is built in Python with FastAPI, and the frontend is a Node.js based app that talks to it over a local development server. A seed script fills the database with development data, including a synthetic dataset that stands in for the real IBM Telco Customer Churn dataset until that file is added. An XGBoost machine learning model is trained on this customer data to predict churn probability, and SHAP is used to explain which factors drove each prediction for a given customer, shown as a feature importance chart on that customer's detail page. The core feature described is a five agent retention pipeline coordinated by an orchestrator. These agents assess churn risk, generate explanations, propose retention strategies, and draft customer communications, each backed by a large language model through a provider agnostic client that can be switched between Google Gemini and Azure OpenAI by changing a single setting, with no code changes required. If no API key is configured, the pipeline still runs, since every agent falls back to a deterministic Turkish language template instead of calling a live model, and the API response marks whether an answer came from the template or from Gemini. The dashboard shows KPI totals, a risk tier breakdown, and a list of top at risk customers, while the customers page supports searching, filtering by risk tier or contract type, sorting, and pagination. Each individual agent can be tested on its own with plain Python objects, without needing a database or a live call to Gemini, and the README includes example commands for running the full pipeline end to end against a seeded and scored customer. Requirements are Python 3.11 or later and Node.js 20 or later, plus an optional Gemini or Azure OpenAI key once agents start calling a real model. The full README is longer than what was shown.