Hasna Khantouch

Designed and implemented core AI tutoring features—including Socratic loop detection, a tiered fallback mechanism, and dynamic questioning modes—within a curriculum-first Python architecture to enhance instructional quality and student engagement. Architected a fallback system to maintain learning momentum, reducing user drop-off rates by effectively detecting and responding to unproductive interactions. Engineered a curriculum-aware tuning layer that adapts examples and terminology for international educational standards without compromising the AI's core Socratic teaching methodology.

Designed and deployed a hybrid security framework for a legal AI chatbot, integrating real-time ML classification and automated evaluation to mitigate adversarial threats while ensuring GDPR compliance. Engineered a two-stage classification system (prompt + attack) with a 33-label taxonomy, reducing vulnerability to novel attacks and maintaining sub-2-second response latency under load. Built end-to-end automated pipelines for model retraining, RAG evaluation, and dashboard monitoring, enhancing system resilience, legal accuracy, and operational transparency.

Built a VS Code extension for automated code auditing using the OpenAI API, leveraging prompt engineering to analyze chunked code in batches and return LLM-based evaluations directly in the extension UI. Designed and integrated a results dashboard for code quality and vulnerability insights, enabling clear feedback and actionable summaries for developers. Implemented real-time token usage tracking and optimization strategies to manage OpenAI GPT-4 API costs efficiently within extension workflows.