BenHassenIlef/csisversion

CSIS is a full-stack AI platform for criminal risk prediction and Tunisian legal assistance. It integrates an XGBoost-based risk prediction engine with a RAG-powered legal chatbot, offering real-time analytics, source-cited legal answers, and an interactive dashboard built with FastAPI and React. # CSIS — Criminal Study Intelligent System AI-powered criminal behavior risk assessment & Tunisian legal intelligence platform --- ## What is CSIS? CSIS (Criminal Study Intelligent System) is a full-stack web platform that combines two AI engines to assist law enforcement professionals and legal researchers in Tunisia. It turns raw socio-behavioral data into actionable criminal risk intelligence and enables natural-language querying of Tunisian law. | Engine | Description | |--------|-------------| | **Risk Predictor** | XGBoost model trained on 10,000 records — predicts criminal behavior risk (%) and likely crime type from 18 socio-behavioral features | | **Legal Chatbot** | RAG pipeline over 1,485+ Tunisian law articles — answers legal questions with cited sources, powered by Groq LLaMA 3.3 / Gemini Flash | --- ## Platform Description ### Risk Assessment Engine The risk assessment module accepts a 18-field profile describing an individual's socio-demographic and behavioral background. It outputs: - **Risk percentage** (0–100%) — how likely this profile leads to criminal behavior - **Binary decision** — "Futur criminel" or "Non criminel" (threshold: 20% probability) - **Predicted crime type** — one of 21 crime categories (e.g. "Vol avec violence", "Trafic de drogue") The prediction runs in real time via a REST API, with results displayed on an interactive dashboard with visual indicators, risk gauges, and history tracking per user session. ### Legal Intelligence Chatbot The chatbot provides natural-language access to Tunisian law. Users type questions in French (or Arabic) and receive direct answers with source citations drawn from actual legal texts. The pipeline works as follows: 1. The user's query is vectorized using a TF-IDF model (bigrams, 1,485+ chunks) 2. Top-k most relevant article chunks are retrieved via cosine similarity 3. The retrieved context + question are sent to the LLM (Groq LLaMA 3.3 70B) 4. The LLM generates a ground …