Selected works

Projects

EURECOM · Sophia Antipolis, France

Meta CRAG-MM Challenge

Multi-modal RAG for Real-time QA · Apr. 2025 – Jul. 2025

PythonPyTorchvLLMTransformersLlama-3RAGVision-Language Models

A multi-modal RAG system for Meta's CRAG-MM Challenge, answering real-time, multi-turn questions over images, knowledge graphs, and web sources with Vision-Language Models.

Key contributions

  • Fused image, knowledge-graph, and web context into a single retrieval pipeline
  • Used Vision-Language Models for multi-modal reasoning over image queries
  • Served Llama-3 via vLLM for efficient, real-time inference

EURECOM · Sophia Antipolis, France

Benchmarking LLMs for Socratic Interactions

Building Educational AI Evaluation Pipelines · Oct. 2024 – Feb. 2025

PythonTransformersPrompt EngineeringNLPLLM Evaluation

An end-to-end pipeline generating Socratic teacher–student dialogues to train educational LLMs, plus experiments to find the best judge model for automated evaluation.

Key contributions

  • Generated Socratic teacher–student conversations as fine-tuning data
  • Benchmarked candidate judge models for automated pedagogical evaluation
  • Packaged reusable tooling for evaluating educational AI systems

KTH Royal Institute of Technology · Stockholm, Sweden

AI in Breast Cancer Mammography

Meta-analysis of AI Diagnostic Performance · Sept. 2023 – Jan. 2024

ResearchMeta-analysisStatisticsMedical AIPython

A co-authored meta-analysis of how reliably AI diagnoses breast cancer from mammograms, synthesizing results across multiple peer-reviewed studies.

Key contributions

  • Ran a systematic review across multiple published studies
  • Performed statistical synthesis to quantify AI diagnostic accuracy