Mohammed Benslimane, Developer in Dubai, United Arab Emirates
Mohammed is available for hire
Hire Mohammed

Mohammed Benslimane

Bio

Mohammed is an artificial intelligence (AI) expert. He previously oversaw AI at one of the cutting-edge startups in deep learning in Europe. After graduating from top institutions, he worked for more than seven years in the conception of machine learning (ML) models, from proof of concept to industrialization in the cloud or embedded devices. Mohammed's work with ML is very expansive and covers real-time recognition of groups of people, satellite imagery, and various systems.

Portfolio

Hudl - Europe
Machine Learning, PyTorch, Python, PyTorch Lightning, Polyaxon...
BeeFutures.io
Edge AI, Hailo, TPU, Coral Services Framework, Raspberry Pi, Transformers...
ENGIE UK & Ireland
Large Language Models (LLMs), Cog-VLM, Optical Character Recognition (OCR)...

Experience

  • Deep Learning - 5 years
  • Machine Learning - 5 years
  • Artificial Intelligence (AI) - 5 years
  • Predictive Modeling - 4 years
  • Computer Vision - 4 years
  • Recommendation Systems - 3 years
  • PyTorch - 3 years
  • Natural Language Processing (NLP) - 2 years

Preferred Environment

Ubuntu Linux, GitHub, Docker, Kubernetes, Visual Studio Code (VS Code), Amazon Web Services (AWS), TensorFlow, PyTorch, Slack

The most amazing...

...project I've worked on is a recognition system engine in crowded areas deployed on 1,000 cameras that is RGPD compliant and attested as a leading solution.

Work Experience

ML Research Engineer

2024 - 2025
Hudl - Europe
  • Led the consolidation of over 40 fragmented ML repositories into a unified, multi-task framework, enabling the deployment of RL-compatible agents for diverse tasks (object detection, event classification, and re-identification).
  • Defined a generic I/O schema to standardize how ML models interact with data—modeling state-action-reward loops across vision-based systems, laying the foundation for decision agents in live video analysis.
  • Designed a system for event tracking from multi-view sports footage, treating each frame as an observation in a temporal simulation environment, where detection models act as low-level agents (reward estimators and highlight selectors).
  • Engineered a training and validation interface where model outputs could be scored against behavioral objectives—simulating reward feedback such as coverage completeness, replay accuracy, and viewer salience.
  • Designed architecture decoupling policy modules (models) from task environments, enabling future extension to reinforcement-style training via gym-like wrappers or real-time feedback loops.
Technologies: Machine Learning, PyTorch, Python, PyTorch Lightning, Polyaxon, Machine Learning Algorithms, Deep Reinforcement Learning, OpenAI Gym, R

AI Expert

2022 - 2024
BeeFutures.io
  • Built a digital beehive AI stack, integrating computer vision analytics.
  • Developed well-engineered real-time object detection and tracking algorithms with 100% accuracy for crowded bee environments (annotation pipeline, QA pipeline for models, camera calibration pipelines, and training pipelines) at 30 fps on Raspberry.
  • Invented and developed custom layers for small object detection (5-pixel objects) using a state-of-the-art few-shot learning paradigm maintaining the same frame rate (30 fps) with 98% map.
  • SHowed my expertise in AI accelerators for edge computing. Optimized DNNs on Raspberry Pi with Hailo AI accelerators (from 5 to 30 fps), decreasing client costs by around 80%.
  • Developed unsupported operations and layers with quantization methods on TFlite to be supported by Coral TPU.
  • Created a library to streamline deep learning model export, managing layers and operations compatibility.
  • Architected and implemented AI audio services. Leveraged cutting-edge signal processing methodologies to isolate meaningful patterns from noisy datasets due to disparity between different colonies.
  • Built attention models (a variant of the LAS model) tailored to analyze colony behavior and provide early swarm detection with 98% accuracy compared to the 58% accuracy of the state-of-the-art models.
  • Designed scalable pipelines with CTO for audio processing and AI model integration, ensuring robust performance in real-time environments.
Technologies: Edge AI, Hailo, TPU, Coral Services Framework, Raspberry Pi, Transformers, Deep Learning, Digital Signal Processing, Audio Processing, Neural Networks, Computer Vision, Machine Learning Operations (MLOps), TensorFlow, TFlite, PyTorch, MNN, ONNX Runtime, Open Neural Network Exchange (ONNX), Calibration, Object Detection, YOLOv5, YOLOv8, You Only Look Once (YOLO), ML Pipelines

AI Expert

2022 - 2023
ENGIE UK & Ireland
  • Led the end-to-end implementation of a digital twin software.
  • Digitized millions of P&IDs in a PDF format using custom OCR and AI-based symbol recognition algorithms (98% accuracy compared to 78% SOTA models).
  • Developed and deployed P&IDs as a graph database in Neo4j, enabling engineering teams to quickly access critical plant data and streamline operations (10x quicker).
  • Enhanced data discovery capabilities by applying deep graph analysis techniques, uncovering hidden patterns, finding critical nodes, and helping mitigate failures.
  • Accelerated P&ID generation 120x faster using RAG on the T5 model (text-to-text transformer) by converting diagrams to text-based flowsheet representation.
  • Led the AI development of the Enershare European project on wind turbine anomaly detection.
  • Developed and fine-tuned a CoG-VLM (Cognitive Vision-Language Model) to enhance anomaly detection for wind turbines, effectively identifying rare edge cases with limited labeled data.
  • Applied graph-based attention mechanisms and regression head on the Cog-VLM model, improving the detection of subtle anomalies and boosting recall by 38% in real-world scenarios compared to our European competitors.
  • Built a scalable Kubernetes-based architecture with the MLOps team to host the anomaly detection system under OpenAPI Standards, enabling near real-time inference under Enershare service requirements.
Technologies: Large Language Models (LLMs), Cog-VLM, Optical Character Recognition (OCR), Neo4j, Retrieval-augmented Generation (RAG), Kubernetes, OpenAPI, Predictive Maintenance, Natural Language Processing (NLP)

Computer Vision Expert

2022 - 2023
BioSort AS
  • Designed and implemented the computer vision stack from R&D to an ML production pipeline. Contributed to the roadmap and strategy design and planning.
  • Developed an advanced (2D and 3D) and accurate camera calibration tool (< 1-pixel error). Features include intrinsic and extrinsic camera calibration, pose estimation, and depth estimation, among others.
  • Developed the fish ID (identification of each individual fish) based on ad hoc deep learning landmarks through various stereo sensors (3D object detection, tracking, and reidentification on highly occluded and crowded fish scenes) in real-time.
Technologies: Computer Vision, C++, OpenCV, Deep Learning, Python, PyTorch, Object Tracking, Geometry, Apache Spark, Data Analysis, Statistical Analysis, Object Detection, Node.js, Internet of Things (IoT), AI Design

Software Engineer

2021 - 2022
Tesla
  • Developed and deployed a pre-production-ready real-time 3D long-range segmentation and detection software for Tesla's autonomous driving system, capable of accurately identifying objects up to 300 meters ahead (98% MAP).
  • Leveraged stereoscopic camera systems to achieve precise depth estimation, enabling high-fidelity 3D scene reconstruction in dynamic driving environments.
  • Applied state-of-the-art 3D techniques such as VoxelNet for point cloud processing and scene segmentation achieving >95% detection accuracy across diverse environments (urban, highway, and mixed weather) within latency constraints.
  • Engineered robust software pipelines utilizing SLAM (Simultaneous Localization and Mapping) and multi-sensor fusion techniques to enhance situational awareness and improve detection reliability at long ranges (up to 300m).
  • Spearheaded model optimization efforts, leveraging techniques like quantization, pruning, CUDA development, and deployment-friendly frameworks such as TensorRT to meet Tesla's stringent real-time performance benchmarks (3 fps).
  • Collaborated with cross-functional teams, including hardware engineers, to integrate the service seamlessly into Tesla’s Full Self-Driving (FSD) stack.
Technologies: OpenCV, Open3D, Object Detection, Image Segmentation, Point Clouds, Simultaneous Localization & Mapping (SLAM), Quantization, Neural Network Pruning, NVIDIA CUDA

Lead Machine Learning Engineer

2021 - 2022
Norbert Health
  • Led the Norberth Health engineering machine learning team in the design, implementation, deployment, monitoring, and testing of several core features in their embedded device product, a Jetson Nano processor.
  • Developed a remote photoplethysmography (rPPG) algorithm for accurate heart rate and heart rate variability (HR and HRV) with 95% accuracy completed compared to oximeter values.
  • Developed a recommendation system based on vital signs and patient metadata for optimal health monitoring. Monitored areas included diet, sleeping hours, doctor appointments, and health insurance.
  • Developed and led the shipping of a proprietary neural network architecture convenient for thermal (IR) and RGB sensors for patient recognition. Increased accuracy by 20% and ten times more FPS than the old algorithm (FaceNet and Chinese whispers.).
  • Designed and implemented large-scale pub/sub message queues using Apache Kafka streams to process the IoT metadata and create Kafka topics for application and system logs.
Technologies: Deep Learning, Computer Vision, Docker, Management, Recommendation Systems, Machine Learning, PyTorch, TensorFlow, NVIDIA CUDA, TensorFlow Deep Learning Library (TFLearn), Bayesian Inference & Modeling, Probabilistic Graphical Models, Time Series Analysis, Image Recognition, OpenCV, NumPy, Point Clouds, MySQL, Data Analysis, Statistical Analysis, Object Detection, Internet of Things (IoT), Optical Character Recognition (OCR), AI Design

ML Software Engineer

2020 - 2021
Meta
  • Designed, trained, and evaluated advanced Deep RL agents in custom high-dimensional environments for strategic learning and control.
  • Designed and compared policy-based and value-based RL architectures, including PPO and A2C for continuous control tasks with dense and sparse rewards and DQN variants (e.g., Double DQN and Dueling DQN) for discrete planning tasks.
  • Conducted extensive benchmarking using Stable Baselines3, RLlib, and custom PyTorch models. Tuned hyperparameters via Optuna and analyzed learning dynamics via Wandb dashboards.
  • Built and maintained end-to-end ML pipelines on PyTorch, ensuring seamless integration with Meta’s internal infrastructure for real-time predictions.
  • Led the implementation of differential privacy techniques to guarantee compliance with user data privacy regulations.
Technologies: PyTorch, Transformers, Recommendation Systems, Machine-learned Ranking (MLR), Differential Privacy, Deep Reinforcement Learning

Lead Data Scientist

2019 - 2020
Servier
  • Led the development of AI methodologies to identify drug targets for pancreatic cancer, overseeing technical innovation and cross-functional collaboration.
  • Defined and implemented roadmaps for cross-institutional collaboration, streamlining data sharing and model development processes in GCP, ensuring alignment with regulatory and data privacy standards.
  • Designed and optimized Graph Neural Networks (GNNs) like GAT and MPNN, improving target prediction precision by 15%.
  • Built AI-powered recommendation algorithms, identifying five novel drug targets and initiating pre-clinical validation.
Technologies: Google Cloud Platform (GCP), Algorithms, Graph Neural Networks (GNNs), Neo4j

Head of Artificial Intelligence

2016 - 2019
Digeiz
  • Developed cutting-edge computer vision solutions for real-time multi-camera and multi-object tracking, driving innovation and deployment in the retail industry.
  • Achieved 98% mAP in real-time multi-camera reidentification (MTMC for CCTV), verified by Bureau Veritas in Paris, outperforming competition through a weakly supervised domain adaptation approach and clustering via the dominant set.
  • Deployed MTMC solutions across three commercial malls, managing 300 cameras per mall, ensuring seamless performance and high scalability.
  • Designed and deployed real-time multi-object tracking systems in highly crowded environments (200+ people per scene), leveraging CUDA for efficient graph-based optimization using the maximum weighted independent set problem.
  • Implemented object detection pipelines for visitor demographics (gender and age) using fisheye cameras, handling 50 cameras per GV100 GPU server, providing actionable insights for retail innovation (99% accuracy).
  • Hired and led a team of four R&D engineers, defined scientific roadmaps, and supported C++ production architecture and AI MLOps pipelines for research collaboration.
Technologies: Deep Learning, Data Mining, Computer Vision, NVIDIA CUDA, C++, Python, Clustering, Anomaly Detection, Pricing, TensorFlow, PyTorch, NVIDIA TensorRT, Siamese Neural Networks, Convolutional Neural Networks (CNNs), Directed Acrylic Graphs (DAG), Neural Networks, Point Clouds, Camera API, Depth Estimation, Image Recognition, Algorithms, Pandas, SQL, MySQL, Data Analysis, Statistical Analysis, ETL, Object Detection, Internet of Things (IoT), AI Design

Data Scientist

2015 - 2016
Adot
  • Built real-time bidding algorithms based on contextual reinforcement learning, including a 30% cost optimization for the same CTR.
  • Reduced the training time of a model used as a prior on ad space value segmentation by 90% while integrating it with Hyperopt and MLflow, which enabled us to perform hyperparameter tuning at scale and deliver value faster.
  • Created and produced a machine learning model from scratch capable of detecting users on different devices (smartphones, laptops, desktops, and tablets), which led the company to change its marketing positioning.
Technologies: Reinforcement Learning, Real-time Bidding (RTB), Pricing, Clustering, Spark, Scala, Machine Learning, Apache Spark, Data Analysis, Statistical Analysis, ETL

Data Scientist

2015 - 2016
Kayrros
  • Participated in a $9 million Series A fundraising by contributing to the development of the Kayrros Oil Index, a cutting-edge solution for weekly tracking of oil production and storage volumes.
  • Processed satellite images to detect spatiotemporal morphological changes in oil basin shapes, providing key insights into storage capacity and production trends.
  • Implemented topic detection algorithms on commodity reports scraped weekly, extracting information for analytics.
  • Designed and built robust databases in Elasticsearch to efficiently store and retrieve weekly processed unstructured data.
  • Developed predictive models using LSTMs to forecast weekly oil production and storage volumes by integrating multi-modal data sources.
  • Collaborated with stakeholders to define and integrate KPIs related to oil production and storage into the Kayrros analytical dashboard, ensuring actionable insights for end-users.
Technologies: Satellite Images, GIS, PySpark, Optical Character Recognition (OCR), Apache Kafka, Elasticsearch, Computer Vision, Predictive Modeling, Behavioral Economics, Docker, Git, Long Short-term Memory (LSTM), Convolutional Neural Networks (CNNs), Keras, Apache Spark, Data Analysis, Statistical Analysis, ETL, Object Detection, Multimodal Models

Experience

Conversational AI for Healthcare Chatbots

WORK DONE
• Built a domain-specific conversational AI system using Transformer-based dialogue models like DialoGPT and RAG (retrieval-augmented generation).
• Integrated the model with knowledge retrieval systems for real-time, contextually relevant responses, achieving state-of-the-art intent accuracy.

Reinforcement Learning in Algorithmic Trading

Machine learning is a relatively new approach to optimizing a trader's portfolio in financial trading. In this sense, predicting the rise or the fall of asset prices is good leverage that can be exploited in buying an asset before the price rises and shorted before it declines. Subsequently, we used reinforcement learning methods to optimize the allocation between risky and riskless assets in this project. Sending signals and receiving rewards from the environment is the key to any reinforcement learning algorithm. So for this purpose, we presented Q-learning to backtest the strategy in terms of cumulative profits by maximizing the Sharpe ratio's value function. Hence, a recurrent reinforcement learning (RRL) algorithm was deduced from this strategy. Simple examples based on a method described above and tested on ”real world” demonstrated promising results and a Sharpe ratio of four or above.

Robust Tracking Using a Collaborative Model

This project is a robust object tracking algorithm using a collaborative model as the main challenge for object tracking is to account for drastic changes in appearance. I proposed a robust appearance model that exploits holistic templates and local representations.

I developed a sparsity-based discriminative classifier (SDC) and a sparsity-based generative model (SGM). Results outperformed state-of-the-art technology (MOT challenge) in complex scenes.

Education

2015 - 2016

Master's Degree in Mathematics, Vision, and Machine Learning

Ecole Normale Supérieure - Paris, France

2012 - 2016

Engineer's Degree in Informatics and Applied Mathematics

Ecole Centrale Paris - Paris, France

2014 - 2015

Master's Degree in Economics

ETH Zurich - Zurich, Switzerland

Skills

Libraries/APIs

TensorFlow, PyTorch, TensorFlow Deep Learning Library (TFLearn), NumPy, Pandas, Scikit-learn, Keras, PySpark, Camera API, OpenCV, Node.js, Open3D, OpenAPI, PyTorch Lightning

Tools

GitHub, MATLAB, GIS, Git, Slack, Open Neural Network Exchange (ONNX), You Only Look Once (YOLO), ChatGPT, OpenAI Gym

Languages

Python, SQL, C++, Scala, R

Paradigms

Anomaly Detection, Siamese Neural Networks, Management, ETL, Machine-learned Ranking (MLR)

Platforms

NVIDIA CUDA, Ubuntu Linux, Docker, Apache Kafka, Amazon Web Services (AWS), Kubernetes, Visual Studio Code (VS Code), Google Cloud Platform (GCP), Raspberry Pi, Polyaxon

Storage

MySQL, Amazon S3 (AWS S3), Elasticsearch, Neo4j

Frameworks

Spark, Apache Spark, Coral Services Framework

Other

Data Mining, Machine Learning, Deep Learning, Artificial Intelligence (AI), Computer Vision, Recommendation Systems, Optical Character Recognition (OCR), Predictive Modeling, Real-time Bidding (RTB), NVIDIA TensorRT, Convolutional Neural Networks (CNNs), Directed Acrylic Graphs (DAG), Image Recognition, Classification, Neural Networks, Time Series Analysis, Algorithms, Mathematics, Data Science, Statistics, Data Analysis, Statistical Analysis, Object Detection, Internet of Things (IoT), AI Design, Forecasting, Applied Mathematics, Natural Language Processing (NLP), Finance, Pricing, Data Visualization, Satellite Images, Long Short-term Memory (LSTM), Time Series, Healthcare Services, Autonomous Navigation, Point Clouds, Generative Pre-trained Transformers (GPT), Global Project Management, Reinforcement Learning, Econometrics, Behavioral Economics, Remote Photoplethysmography (rPPG), Clustering, Deep Reinforcement Learning, Trading, Bayesian Inference & Modeling, Probabilistic Graphical Models, Depth Estimation, Health, Biometrics, ML Pipelines, Object Tracking, Geometry, Multimodal Models, Transformers, Differential Privacy, Graph Neural Networks (GNNs), Image Segmentation, Simultaneous Localization & Mapping (SLAM), Quantization, Neural Network Pruning, Large Language Models (LLMs), Cog-VLM, Retrieval-augmented Generation (RAG), Predictive Maintenance, Edge AI, Hailo, TPU, Digital Signal Processing, Audio Processing, Machine Learning Operations (MLOps), TFlite, MNN, ONNX Runtime, Calibration, YOLOv5, YOLOv8, Chatbots, knowledge retrieval systems, DialoGPT, Machine Learning Algorithms

Collaboration That Works

How to Work with Toptal

Toptal matches you directly with global industry experts from our network in hours—not weeks or months.

1

Share your needs

Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2

Choose your talent

Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3

Start your risk-free talent trial

Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.

Top talent is in high demand.

Start hiring