Experience

A summary of my professional roles and educational background. I thrive on tackling challenging problems and continuously learning new skills.

Professional Experience

Extracted named entities (metadata, paradata) from unstructured cultural heritage analysis data using LLMs, Prompting, and RAG to ensure provenance and traceability.
Optimized named entity extraction through the implementation of LLM reasoning techniques (Chain-of-Thought, internal reasoning).
Fine‑tuned Whisper models with LoRA on cultural heritage audio data, enhancing transcription accuracy of domain‑specific named entities by 5%.
Linked extracted entities to Opentheso and ORCID knowledge bases, strengthening data accessibility and traceability.
Partnered with heritage science experts (documentalists, conservators, restorers) to create training and evaluation datasets.

Automated the release of new versions of the PARSEME corpora across 26 languages via GitLab CI/CD, considerably reducing publication times.
Designed and implemented new modules for quality control of linguistic corpora.
Unified and integrated PARSEME’s disparate tools, increasing operational efficiency by 50%.
Enhanced French multiword expression identification F1-score by 10% leveraging CamemBERT and FlauBERT models.
Built an inference pipeline on lab clusters to identify multiword expressions in a large 21GB corpus.

Designed and implemented a lightweight, efficient CNN for avionics protocol recognition using Tensorflow, achieving aerospace‑standard \(10\)ms latency and error rate below \(10^{-5}\).
Collaborated with simulation team to enhance protocol recognition through data collection, analysis and production testing.
Orchestrated the deployment of the recognition model on AWS Cloud and Raspberry Pi 4 using TensorFlow Lite, ensuring optimal performance.
Developed and launched an intuitive, interactive dashboard using Dear PyGui for real‑time monitoring of model performance in production.
Developed substitution models using Transformer for dynamic systems.

Programmed Scannol, an AI‑based auditing tool using expert systems to detect anomalies in C/C++ code.
Tracked down and identified efficiently functional errors in each of the 4 source codes analyzed.
Adopted and mastered the demanding DO‑178C development standards.

Project: Fused question and image attention to optimize visual question answering (VQA) using deep learning techniques.

Programming: Python, C/C++, Java, JavaScript, R, HTML
Machine Learning: Tensorflow, PyTorch, Hugging Face, Scikit‑learn, LlamaIndex, OpenCV, spaCy, Kubeflow, Spark, Slurm
Database: SQL, NoSQL, PostgreSQL, Oracle
Exploratory Data Analysis: Seaborn, Matplotlib
Cloud Computing: Docker, Kubernetes, AWS, GCP
Software Development: Git, Jira, Agile, Anaconda, Visual Studio, Django, FastAPI, Jenkins, CI/CD, Bash