Elyes Ben Chaabane

– Designed and benchmarked ML models for disease prediction (280+ conditions) and readmission forecasting, with focus on reproducibility and production-readiness – Improved research codebase quality (60% reduction in training time) and experiment tracking – collaborated with international researchers and professors (Stanford, NUS, NTU, ETH)

– Built reasoning dataset (200K+ samples) by distilling large models (GPT-OSS 120B); compared fine-tuning strategies (LORA/QLoRA/full), achieving 15% improvement in reasoning accuracy – Extended HuggingFace TRL library to implement Magistral's GRPO; designed a custom reward models

– Lead technical strategy for client projects delivering web/mobile apps and ML solutions (LLM-based systems, text classification, multimodal agents) – Managed cross-functional teams of 3-4 members per project; owned project scoping, budgeting, and client communication

– Developed and benchmarked deep learning architectures for rare genetic disease detection from craniofacial images, improving the accuracy by 10% on a dataset of 5,000 patients – Built explainable AI pipeline for clinician validation; partnered with surgeons and CTO to balance clinical and technical constraints

– Advised on technology stack and memory architecture for an AI agent platform using open-source frameworks – Provided budget estimation for project development and model deployment

– Coordinated and designed a software development course for 18 students aged between 35 to 50. – The course covered basics of Python, theoretical computer science, and machine learning.

– Developed and trained a new GNN for physics movement prediction using an encoder-decoder architecture – Improved the accuracy by 20%; The model outperformed state-of-the-art models.

– Owned technical architecture and delivery of AI-agent platform – architected RAG pipeline (LlamaIndex) and optimized LLM inference costs by 65% through model quantization and caching – Led multidisciplinary team of 15 engineers (ML, backend, frontend), overseeing sprint planning and delivery; shipped 20+ features across 8 sprint cycles

– Developed, designed, and maintained a comprehensive mobile app for helicopter pilots, facilitating pre-flight checks and in-flight crisis management. – Developed and designed a mobile app with Bluetooth communication capabilities for a multi-sensor temperature measurement device. – Designed a website for KruSwiss

– Trained and compared multiple AI models on markerless pose estimation to automatically measure the Weber's length of an ant. – The model generalized well to unseen ant species

– Generated a QA dataset using GPT-4 and a RAG pipeline on UK's highway code – Used the dataset to fine-tune an LLM and tested it for prompt injections

– Implemented an IoT algorithm to analyze real-time data for the Payload team (article) – Co-authored a paper on liquid slushing dampening (paper)