Mohamed Ali Ben Alaya

Master's Degree in Data Science

2019 - 2021

ESPRIT School of Engineering , Tunisia

During my Master's degree, I specialized in Data Science, being selected among the top 90 out of 500 applicants and graduating in the top 1% of my class with honors.

The program emphasized both advanced mathematics, including probability, algebra, and optimization, and practical development through real-world machine learning and deep learning projects.

Our university had a strong project-based learning approach, where I worked on numerous projects involving Big Data technologies and Generative AI.

For my Master's thesis, I completed a six-month internship at a company in Belgium, achieving a final grade of 18.5/20 and graduating with excellent honors.

Bachelor's Degree in Computer Science

2016 - 2019

ESPRIT School of Engineering , Tunisia

During my Bachelor's degree, I completed three years of preparatory coursework and one year of intensive core training, specializing in programming languages for web, desktop, and mobile application development.

AI Engineer

Feb 2022 - Present

Rutilea, Kyoto, Japan

• Battery CT Defect Detection: Developed a multi-stage AI pipeline for micro-crack detection in battery CT scan frames. Integrated advanced image preprocessing with unsupervised (PatchCore, DRAEM) and supervised (RF-DETR) models. Tuned ensemble strategies and thresholds to handle noisy frames, achieving 98% precision in crack localization under varying scan conditions.
• Voice-to-Video Synthesis: Engineered a voice-to-video synthesis system using SadTalker and a custom-trained Style-BERT-ViT2 model. Focused on realistic facial animation, speech synchronization, and keypoint smoothing techniques to generate natural-looking talking avatars from voice inputs.
• RAG-based Dialogue Simulation: Built a Retrieval-Augmented Generation (RAG) pipeline with LLaMA 3 and LlamaIndex to simulate child-like dialogue behavior. Designed retrieval and generation mechanisms to create contextually relevant and natural responses, including topic classification for downstream tasks.
• Handwritten Receipt OCR: Developed a full OCR pipeline combining YOLO object detection for field localization and fine-tuned TrOCR for handwritten text recognition. Optimized model robustness to noisy handwriting and diverse receipt templates, achieving 95% information extraction accuracy.
• Data Processing APIs: Built high-concurrency FastAPI endpoints and optimized SQL pipelines to manage a 40-table relational database. Designed for efficient statistical aggregation and querying, supporting over 10,000 concurrent users across business applications.
• Multimodal Vision-Language Application: Integrated Florence, PaliGemma, and Donut vision-language models into a multimodal inference system. Deployed a datacenter-hosted service for image-to-text tasks such as captioning, document extraction, and multimodal classification.
• 2D CAD Drawing Retrieval System: Implemented a Vision Transformer (ViT)-based feature extraction pipeline for 2D car part drawings. Indexed extracted embeddings into Elasticsearch using cosine similarity, achieving 85% top-5 retrieval accuracy for production-level part matching and retrieval tasks.

Data Scientist

Oct 2021 - 2022

VALUE,The Banks of the Lake,Tunisia

• Stock Market Prediction Model: Developed a predictive trading model for the Tunisian stock market by integrating Financial BERT embeddings, deep autoencoders, and a range of technical indicators. The model forecasted market trends and optimized trading strategies, leveraging both structured financial data and sentiment analysis from financial news sources.
• Real-Time Trading Dashboard: Designed and deployed an interactive real-time dashboard that visualized model-driven buy/sell signals. The dashboard provided actionable trading insights through dynamic charts, alert systems, and performance tracking, enabling data-driven decision making for traders.

AI Engineer

Mar 2021 - Aug 2021

Becode,Brussels,Belgium

• Yoga Pose Estimation System: Developed a deep learning-based system for real-time yoga pose analysis by combining human pose estimation models Mediapipe,OpenPose with pose classification architectures. Focused on accurate keypoint detection, pose labeling, and real-time feedback for user alignment correction.
• Pose Evaluation Pipelines: Implemented triangulation techniques and keypoint-based evaluation pipelines to assess practitioner posture quality across sequential video frames, enabling precise analysis of body alignment and movement flow.
• Data Collection and Preprocessing: Built scalable pipelines for collecting, labeling, and preprocessing large volumes of pose video data, supporting robust model training and validation for human pose estimation tasks.