Youcef KADDOUR
Lead AI Engineer / Tech Lead
I build production-grade LLM & CV applications that solve real-world problems at scale.
Technical Skills
Technologies and tools I work with
Education & Certifications
Academic foundation and professional certifications in AI, machine learning, and cloud technologies.
Education
Mobile Autonomous Systems
Paris Saclay University
GPA: 4.0
Key Coursework:
Aerospace Telecommunication
Institute Of Aeronautics And Space Studies
GPA: 3.8
Key Coursework:
Aerospace Telecommunication
Institute Of Aeronautics And Space Studies
GPA: 3.8
Key Coursework:
Professional Experience
My journey in AI and machine learning
- Acted as AI consultant and technical expert across aerospace, automotive, industrial, and tech sectors.
- Delivered end-to-end AI/ML solutions including data pipelines, model development, and system integration.
- Led discovery workshops to identify pain points, shape AI strategies, and support pre-sales proposals.
- Produced solution architectures and technical documentation, improving proposal success rates.
- Led R&D initiatives, set technical roadmaps, and mentored engineers and researchers.
- Advised on LLMs, generative AI, computer vision, and ML best practices across client engagements.
Technologies Used:
- Led a 14-person team to build and deploy an industrial AI-agent framework for autonomous production and maintenance workflows.
- Designed a modular multi-agent architecture (planning, control, anomaly detection, explainability) using model-based RL and rule-based reasoning.
- Delivered internal trainings on AI, LLMs, and agentic architectures to upskill engineering teams.
- Built a full MLOps stack (Docker, K8s, MLflow, GitHub Actions) enabling weekly retraining and zero-downtime deployment.
- Established engineering standards and mentorship programs, increasing delivery velocity by 25%.
Technologies Used:
- Led development of a Text-to-SQL system achieving 91% EM accuracy for Airbus Helicopters.
- Built a low-GPU LLM fine-tuning & inference framework (LoRA, Triton, CUDA) reducing GPU usage by 60%.
- Defined multi-criteria evaluation metrics and built the full evaluation pipeline.
- Standardized LLM optimization practices, documentation, and cross-team knowledge transfer.
Technologies Used:
Featured Projects
Open source projects and research work
A multi-tenant AI access platform with governance, budgeting, observability, and secure API key management.
A Python library for efficient LLM fine-tuning with adaptive GPU optimization, 4-bit loading, and LoRA enhancements.
High-performance Transformer inference using Python gRPC with a C++/CUDA backend.
Full-stack AI-powered medical training simulator for students and professors.
An advanced ChatGPT-style chatbot with dual views and custom content blocks, plus an interactive mindmap for exploring conversation flows.
Production-ready RAG system for document Q&A
Weather station using ESP32-S3, 4" RGB touchscreen (ST7701 + GT911), and BME280 sensor with LVGL UI and built-in Wi-Fi.
Autonomous landing system for a Parrot AR Drone 2.0 on a moving TurtleBot platform using visual tracking, IMU fusion, Kalman filtering, and PID control.
AI Models
Custom trained models and research contributions
Quantized LLM for Fast Inference
Llama 3 (8B parameters) quantized to 4-bit using BitsAndBytes
Training Data
Training Setup
Performance Metrics
Limitations
- Quantization reduces precision in long-context reasoning
- Not suitable for high-risk tasks requiring exact token outputs
- Dependent on BitsAndBytes compatibility with hardware
Ethical Considerations
- Follow Meta’s license for any commercial or derivative use
- Must not be used to generate harmful or unauthorized content
- Users should ensure quantized outputs remain safe and accurate
Quantized LLM for Low-Resource Inference
Llama 2 (7B/13B variant) quantized to 4-bit using BitsAndBytes
Training Data
Training Setup
Performance Metrics
Limitations
- Loss of precision in arithmetic and reasoning tasks
- May hallucinate more under long prompts
- Quantized models may be unstable on older GPUs
Ethical Considerations
- Any downstream use must comply with Meta’s Llama 2 License
- Should not be used for safety-critical or regulated decision-making
- Quantized variants must be evaluated before deployment
Text-to-SQL Translation (LoRA Fine-Tuned)
LLaMA 7B (LoRA Low-Rank Adapters)
Training Data
Training Setup
Performance Metrics
Limitations
- Requires careful prompt engineering
- May hallucinate table/column names
- Still struggles with long schema contexts
Ethical Considerations
- SQL outputs must pass validation before execution
- Aviation operational data must be kept private
- Should not be used for mission-critical queries without review
Text-to-SQL Translation (LoRA Fine-Tuned)
Mistral 7B (Dense Transformer, LoRA)
Training Data
Training Setup
Performance Metrics
Limitations
- Still dependent on schema clarity
- Performance varies on extremely long SQL trees
- Fails on rare domain-specific operators
Ethical Considerations
- Human verification required for SQL execution
- Training logs must be anonymized
- High accuracy does not eliminate hallucination risks
Publications & Talks
Research papers, blog posts, and conference presentations
Featured Publications
Development of a real-time detection, monitoring, and landing system enabling an autonomous drone to land on a moving platform. Work included updating and extending a Python library for Parrot AR-Drone 2.0 control, implementing a computer-vision detection pipeline using ArUco markers, Kalman filtering, and GOTURN tracking, and creating a real-time trajectory planning and correction system.
Comprehensive study and categorization of state-of-the-art visual tracking and monitoring methods based on Deep Learning, including CNN, RNN, SNN, YOLO-v3, and DRLT. The project included performance evaluations across multiple benchmarks with respect to robustness, computational requirements, memory usage, and adaptability across various scenarios. Key limitations of each method were identified to support method selection for real-world applications.
Get In Touch
Let's discuss opportunities and collaborations
Blog
Coming soon...