Ali Mokhfi

Robotics and AI systems engineer with end-to-end ownership across the full robotics stack: motor and sensor integration, ROS2 node architecture, motion planning on real 6-DOF hardware, hybrid Visual SLAM pipelines, LiDAR-camera-IMU sensor fusion, sim-to-real validated RL policies, and distributed multi-agent control systems deployed on live industrial hardware. Specializing in autonomous UAV systems, multi-modal perception, and neural world modeling. Experienced in designing and integrating real-world AI systems combining computer vision, control systems, sensor fusion, and industrial automation.

Work focuses on building Neural SLAM–based world models for GNSS-denied environments, enabling unified spatial understanding for navigation, mapping, and high-level inference tasks such as early agricultural disease detection and multi-agent coordination. Strong background in industrial AI applied research within petrochemical environments, where contributed to system-level optimization, stakeholder communication of technical AI systems, and operational efficiency improvements.

Founding R&D Engineer — Robotics & AI Systems - Sonatrach Petrochemical Facility - Skikda, Algeria

(2023-01 - 2024-12)

Full-stack hardware deployment: End-to-end ownership of two AI and robotics system deployments on live industrial DCS infrastructure — from sensor integration and data pipeline design through control software commissioning. Collaborated directly with mechanical, electrical, and plant operations engineers.

• 6-agent MADDPG motor/actuator control: Designed and deployed 6 cooperative MADDPG agents (CTDE, Transformer inter-agent attention) for real-time MIMO distillation column actuation on live industrial hardware. Built perception-to-action pipeline with millisecond latency inference.

Outcome: +4.7% white spirit recovery · +53.4 T/D throughput · ~$180K/year measured value; zero off-spec cycles post-deployment.

• Sensor data pipeline: Integrated real-hardware sensor pipelines over 500K+ live DCS timestamped records across 32 concurrent streams — real-time Kafka ingestion → PySpark preprocessing (timestamp alignment, noise filtering, regime segmentation) → Delta Lake Medallion Architecture → training dataset export.
• Automated fault diagnosis system (RE-DTFN): 3-branch parallel CNN (time-domain + FFT + STFT), cross-domain attention fusion + DQN RL fusion agent for dynamic branch weighting — deployed on live gas-processing equipment.

Outcome: 22% reduction in unplanned shutdowns; 94.3% fault classification accuracy on held-out test set.

• Automated MLOps & inference service: Airflow DAG → Great Expectations → ZenML → MLflow → GitHub Actions → Kubernetes EKS. Evidently AI drift monitoring with automated retraining trigger; p99 latency 7.4 ms; model degradation detection reduced from days to under 2 hours; zero-downtime rollback.
• Commissioning-standard documentation: Produced failure mode analyses, test reports, and system documentation for cross-functional engineering review.

PhD Candidate — Intelligent Systems - Harbin Institute of Technology (HIT) - Shenzhen, China

(2025-01)

• Research: Non-Stationary Behavioral Policy Learning for Autonomous Driving via Latent Dynamic Social Fields — modelling implicit interaction dynamics between autonomous vehicles as latent social fields, enabling policies to adapt to non-stationary agent populations (varying densities, unseen behaviours) without full environment re-training.

PhD - Intelligent Systems - Harbin Institute of Technology (HIT) (2025-01)

MSc - Automation & Robotics Engineering - University of 20 Août 1955 (2023-01 - 2025-12)

BSc - Automation & Control Engineering - University of 20 Août 1955 (2020-01 - 2023-12)