hajar El mahjouby

Conception, développement et déploiement de solutions d’automatisation basées sur l’intelligence artificielle pour optimiser les processus internes et les fonctions support. Mise en place de workflows automatisés, d’assistants IA et de solutions utilisant les LLM, le RAG et les agents IA. Intégration avec les outils métiers via Python, API REST, webhooks, SQL et des plateformes d’automatisation telles que n8n.

Designed and built VoxAdria end-to-end: a voice-to-voice banking assistant supporting Modern Standard Arabic, Moroccan Darija, French and English, covering 4 Moroccan banks (CDM, CIH, Assafa, Attijari), each with its own knowledge base of hundreds of institutional documents. Live-demoed regularly to stakeholders. • Cross-lingual RAG pipeline: queries auto-translated to French, retrieved over French-only banking docs via BGE-M3 embeddings indexed in Qdrant, with grounded answers generated back in the user’s original language. • Web scraping pipeline (Scrapling + BeautifulSoup) for document collection; benchmarked 8 multimodal OCR models on banking layouts (tables, forms, mixed scripts) before selecting LightOnOCR. • Hybrid voice stack with Unicode-based TTS routing: VoxCPM2 via vLLM-Omni for FR/EN, Gemini Flash TTS for Arabic/Darija (VoxCPM under-performed on Moroccan Arabic), Whisper-large-v3-turbo for STT, with SHA-256 disk caching. • Self-hosted multi-LLM serving on dedicated GPU via Ollama (Gemma3-27B, GPT-OSS-20B, Mixtral 8x7B) with models kept resident in GPU memory, exposed through a FastAPI backend (/chat, /tts, /stt) and a voice-interactive React frontend.

• Automated document workflows using n8n • Processed files & created OpenAI embeddings • Stored vectors in Supabase, metadata in PostgreSQL • Built a RAG-based chat interface for smart document search • Integrated with FTP, scalable via Docker • Hands-on with LLMs, NLP, and containerization