Kiprotich Dominic

EV-Failed-Session Resolver diagnoses failed EV charging sessions from three inputs: a driver complaint (8 languages supported), an OCPP-style charger log, and a mobile-money payment record. It returns a cited diagnosis with verified corpus references, a driver reply in the complaint's language, and a refund draft that always requires human approval. The core design rule: the model (Muse Spark) only reads messy text and writes narrative with placeholders all arithmetic, citation verification, and escalation decisions happen in audited Python code. Every run is written to a SQLite audit log with its prompt, raw model output, and validation result, and an evaluation harness compares the model against a code-only baseline on human-written labels. Built with FastAPI, Next.js, and Muse Spark. All demo data is synthetic. MIT licensed; an independent project.