HayderCH/multilingual-fleet-ops-assistant

Multilingual AI assistant for fleet operations, supporting French, Tunisian Arabic and Arabizi through safe intent routing, structured API workflows and evaluated fallback mechanisms. # Multilingual Fleet Operations Assistant (AI Engine) An end-to-end, **100% self-contained local AI assistant** for fleet operations. It processes multilingual natural language queries across **French, Tunisian Arabic (Derja), Arabizi (Latin script with phonetic numerals), and English**, resolves intent and parameters, manages multi-turn clarification and confirmation state, and executes structured tool calls against fleet backends. > [!NOTE] > **Portfolio & Clean-Room Provenance Notice** > This repository is an open-source reference implementation written from scratch with synthetic data and mock adapters. It contains no proprietary client source code, credentials, or private vehicle telemetry. --- ## 📸 Interface Preview Interactive local web console demonstrating multi-turn clarification, Arabizi parsing, and inspectable routing decisions (intent, status, confidence score). --- ## ⚡ Key AI Engineering Highlights * **Multilingual & Arabizi Normalization:** Normalizes code-switched queries across Arabic script, French, English, and Arabizi (phonetic numbers like `3` for ع, `7` for ح, `9` for ق). * **Agentic Guardrails & Tool Calling:** Strict schema validation isolates natural language understanding from execution; the assistant only triggers allowlisted operations with validated parameters. * **Multi-Turn State & Clarification Loops:** Automatically identifies missing parameters (e.g. missing vehicle ID) and prompts the user for clarification before executing backend calls. * **Confirmation Gates on State Mutations:** Enforces explicit user confirmation before executing state-changing actions (e.g., ticket/complaint creation). * **Self-Contained & Reproducible:** Zero required cloud API keys or external paid LLM dependencies; runs entirely locally using FastAPI, scikit-learn / n-gram intent classification, and Redis-backed session memory. --- ## 🏗️ Architecture & Data Flow ```mermaid flowchart TD subgraph Client Layer U[User Query in FR / TN / …