Millie-10/phelix

Fraud intelligence platform for African fintech - ML + LLM cascade architecture # Phelix — Fraud Intelligence Platform > Named after Felix. Built to protect people's money. Phelix is an AI-powered fraud detection system that classifies customer messages in real time, routes them to the right team, and escalates edge cases to a large language model when confidence is low. --- ## The Problem Fraud costs African fintechs billions annually. When a customer contacts their bank about a suspicious transaction, a phishing attempt, or an account takeover — every second matters. Manual routing is slow, inconsistent, and expensive. Phelix solves this by reading the customer's message and instantly classifying the intent — so the right team gets the right case at the right time. --- ## What Phelix Does A customer sends a message. Phelix reads it and returns: - **Intent** — what the customer is actually reporting - **Confidence score** — how certain the model is - **Recommended action** — what the bank should do next - **Method** — whether ML or LLM handled it ```json { "intent": "phishing_attempt", "confidence": 73.6, "recommended_action": "Warn customer and log incident", "method": "ml_model" } ``` --- ## The 7 Intent Categories | Intent | Description | Action | | --------------------- | --------------------------------------- | ------------------------------ | | `legitimate_inquiry` | Normal customer service question | Route to customer service | | `suspicious_transfer` | Urgent or unusual transfer request | Hold and flag for review | | `account_takeover` | Someone else accessing the account | Lock account, alert security | | `phishing_attempt` | Scam or social engineering attempt | Warn customer, log incident | | `dispute_claim` | Incorrect charge or billing error | Open dispute ticket | | `unusual_activity` | Transaction inconsistent with behaviour | Flag for fraud analyst | | `account_recovery` …