ismail benalla

Built Airflow pipelines enriching web, app, and CTV sitelists with Gemini and Google Search grounding. Developed BigQuery-to-Qdrant ingestion and multilingual retrieval for Ray, and led a data lineage technology assessment.

Built a configurable multi-agent extraction platform with custom schemas, API/email/file workflows, FastAPI, PostgreSQL, React, Terraform, and AWS/Azure deployment.

Built an AI email agent generating personalized audio summaries with ElevenLabs; delivered using FastAPI, PostgreSQL, React, AWS, Terraform, and GitHub Actions.

Designed a human-in-the-loop feedback system for continuous agent improvement, increasing document extraction accuracy by approximately 18%. Built a modular agent evaluation framework covering 10+ metrics and refactored the data extraction pipeline using LangChain and LlamaIndex.

Built a real-time voice conversion system using PyTorch and ONNX, achieving end-to-end latency below 80 ms. Developed a multi-voice blending module based on the LLVC architecture and improved MOS by 0.3 points over the StreamVC baseline.

Designed and deployed a production RAG system for Arabic and French legal documents using FastAPI, LlamaIndex, PostgreSQL, and Qdrant, reducing average query latency to below 350 ms. Improved retrieval precision by 18–22% using cross-encoder reranking and hybrid search.

Master's in Artificial Intelligence with focus on Deep Learning, NLP, and MLOps from Centrale Lyon, from September 2024 to October 2025.

Computer Science Engineering Degree from the National School of Applied Sciences, Fez, from September 2019 to August 2024.