Earth observation
Super-resolution, restoration and segmentation of Sentinel-1 SAR and Sentinel-2 multispectral imagery, including temporal and cloud-aware modelling.
AI / Machine Learning Engineer
Bilel
Khlaifia
I build and optimise production deep-learning systems: computer vision, geospatial machine learning, and distributed training on NVIDIA DGX, A100 and H100 clusters.
01
What I work on
Distributed training
PyTorch DDP and NCCL across multi-node A100 and H100 clusters, with mixed precision, activation checkpointing and sharded data loaders.
Inference optimisation
Structured pruning, INT8 and FP16 quantisation, ONNX export and TensorRT engine tuning, profiled with Nsight Systems and Compute.
Edge perception
Low-power computer vision on NVIDIA Jetson-class hardware for autonomous platforms, with embedded C++ and custom CUDA kernels.
02
Selected work
Production systems
Commercial work. Source is private; described by capability and stack.
Sentinel-2 super-resolution
Latent-diffusion super-resolution for four-band Sentinel-2 imagery, with 2.5 m and approximately 1.67 m release profiles, served through a tiled inference API.
- PyTorch
- Diffusion
- TensorRT
- FastAPI
Methane & emissions intelligence
Multispectral ingestion and plume-detection pipeline serving GeoJSON and COG endpoints, with GeoServer WMS temporal visualisation behind a map viewer.
- Python
- GeoServer
- COG
- TerriaJS
Crop intelligence model
Multimodal, multitemporal remote-sensing model for crop segmentation and classification, with conceptual lineage from MAESTRO.
- PyTorch
- ViT
- Self-supervised
InSAR ground motion
Sentinel-1 PSI/SBAS rolling pipeline following the DISP-S1 pattern, producing per-cycle ground-motion products over regional areas of interest.
- ISCE3
- COMPASS
- dolphin
Water-level forecasting
Temporal Fusion Transformer producing 240-hour water-level forecasts, served through GPU and CPU inference backends with a dashboard front end.
- TFT
- PyTorch
- FastAPI
- React
On-device 3D measurement
iOS LiDAR depth pipeline for physical measurement: point-cloud processing, geometric calibration and uncertainty estimation with approximately 1 mm end-to-end measurement accuracy.
- ARKit
- LiDAR
- Swift
Open source
Public repositories. Star counts refresh live from the GitHub API.
Jetson-accelerated visual navigation for drones and ground vehicles using TrailNet, stereo depth estimation, ROS and PX4 control.
- C++
- TensorRT
- ROS
- BSD-3-Clause
PyTorch U-TAE model for binary water and land segmentation from Sentinel-1 and Sentinel-2 imagery using the IBM Granite flood dataset.
- Python
- Temporal attention
- GPL-3.0
Real-time webcam object detection with OpenCV DNN and a pre-trained MobileNet SSD model in Caffe format.
- Python
- OpenCV
- MIT
PyTorch CNN for classifying EuroSAT RGB image patches into ten land-cover classes.
- PyTorch
- Land cover
- MIT
Airbus satellite-image sample dataset with aircraft annotations and a YOLOv5 detection experiment.
- YOLOv5
- Detection
- CC BY-NC-SA
AirSim-based code for camera-driven UAV navigation experiments using reinforcement learning.
- Python
- RL
- MIT
03
Technical blog
Sentinel-first crop intelligence: joint 10 m mapping and constrained ×10 detail
A rigorous account of a compact multimodal temporal model that combines Sentinel-1 and Sentinel-2 for crop segmentation, parcel-boundary estimation and model-derived detail on a 1 m storage grid.
- Remote sensing
- Temporal Transformer
- Multi-task learning
- Reproducibility
Scaling geospatial deep learning: from a single GPU to distributed PyTorch training
A measurement-driven study of mixed precision, batch scaling, memory pressure and the evidence required for defensible DDP speedup claims.
- PyTorch DDP
- AMP
- GPU profiling
- Scaling
Beyond PSNR: evaluating satellite super-resolution without high-resolution truth
A scientific audit of observation consistency, spectral drift, learned high-frequency detail and stochastic spread in a Sentinel-2 latent-diffusion release.
- Super-resolution
- Spectral fidelity
- Uncertainty
- Remote sensing
Inside Redtail: how a Jetson flew forest trails, and what preserving that stack required
An engineering audit of NVIDIA’s Redtail autonomy stack — TrailNet orientation estimation, stereo depth networks and PX4/MAVROS control — preserved as a verifiable snapshot with a 2026 modernization blueprint.
- Embedded autonomy
- TensorRT
- ROS
- Software preservation
04
Experience
-
Sep 2021 – Present
Machine Learning Engineer Lead
NextAV · Geospatial AI
- Productionised a system for detecting methane and oil leaks from multispectral satellite imagery, cutting p95 inference latency from 950 ms to 320 ms through structured pruning and TensorRT engine tuning.
- Delivered multi-generation Sentinel-2 super-resolution sustaining more than 1,000 512×512 tiles per hour via mixed precision, fused augmentation and asynchronous I/O.
- Scaled distributed PyTorch training on DGX A100 and Azure H100 clusters to 3.9 times the original throughput using DDP, NCCL, AMP, sharded loaders and activation checkpointing.
- Reduced per-epoch GPU cost by 48% through Nsight profiling, memory reuse and selective precision tuning.
- Built and mentored a hybrid engineering organisation of more than 25 people, instituting profiling standards, reproducible evaluation harnesses and CI/CD quality gates.
-
Sep 2025 – May 2026
Fractional Staff Research Engineer
AtriaGem · Luxury Tech
- Led R&D from concept to prototype, translating research into production-ready systems.
- Built on-device 3D vision for iOS ring sizing using ARKit LiDAR depth, reaching approximately 1 mm end-to-end measurement accuracy.
-
Jun 2024 – Present
AI Research Consultant
Startupbootcamp · Accelerator
- Standardised MLOps and GPU-acceleration templates across more than five portfolio companies, shortening deployment timelines.
- Advised founders through workshops and mentor sessions on technical strategy and go-to-market narrative.
-
Jan 2019 – Feb 2020
Autonomous Systems Research Lead
AVIONAV · Ultralight Aircraft
- Prototyped an autonomous quadcopter integrating Pixhawk PX4 flight control with a Jetson AGX Xavier companion computer for GPS and vision-assisted autonomy.
- Developed an embedded C++/CUDA perception stack performing object detection at 60 fps with a power draw below 15 W.
05
Research & teaching
Publications
-
Few-Shot Crop Mapping with Vision Transformers
ISPRS, 2022 · Oral
Few-shot ViT with temporal token reduction and spectral embeddings; 94.2% macro accuracy using less than 15% of the labels.
-
Multi-Scale Satellite Super-Resolution
Under review, 2026
Cross-scale attention with a spectral-consistency loss for 10 m to 1.5 m reconstruction, with TensorRT-optimised inference.
Open source
-
PyTorch
Build & developer experience
Upstream proposals on editable installs, CMake prefix resolution and Ninja discovery in active Python environments, plus issue reports on CUDA autograd behaviour.
-
GitHub Advisory Database
Open pull requests
Corrections to affected version ranges and package metadata for published security advisories.
Teaching
-
Assistant Professor, AI & Data Science
Military Aviation School of Borj El Amri · 2020–2022
Applied AI and geomatics curriculum for remote sensing and spatial analytics.
-
Assistant Professor, Data Science
National Agronomic Institute of Tunisia · 2021–2022
Precision-agriculture AI: spectral analytics, time-series modelling and yield prediction.
-
Assistant Professor, AI & Data Science
Virtual University of Tunis · 2021–2022
Remote instruction to more than 200 students per year on geospatial modelling and agritech.
Recognition
- 2025World LPG Challenge: Winner for AI methane leak detection
- 2024–25AfricArena Summit: exhibitor and speaker in Cape Town
- 2023GITEX Morocco: geospatial AI for energy
- 2021NVIDIA GTC: Jetson Educators Roundtable
- 2020IADE Tunisia: autonomous UAV navigation
Certifications
- 2021–NVIDIA DLI Certified Instructor & Jetson AI Specialist
- 2021–DeepLearning.AI Ambassador
06
Technical
- Languages
- Python · C/C++ · CUDA C++ · Bash · SQL
- Frameworks
- PyTorch · Hugging Face Transformers · TensorRT · JAX · ONNX Runtime · scikit-learn
- Models
- Vision Transformers · CNNs · GANs · Diffusion · LoRA / PEFT · Few-shot · Self-supervised
- Distributed
- DDP · NCCL · AMP · Activation checkpointing · Sharded loaders · Slurm
- GPU
- CUDA kernels · Streams · cuDNN · cuBLAS · Nsight Systems / Compute · Triton Inference Server
- Optimisation
- Structured pruning · INT8 / FP16 quantisation · ONNX export · Distillation · Latency profiling
- MLOps
- Docker · Kubernetes · FastAPI · gRPC · MLflow · Weights & Biases · GitHub Actions · DVC
- Cloud
- AWS (SageMaker, S3, EC2) · Azure ML · Slurm · NVIDIA DGX · H100 / A100
- Geospatial
- Sentinel-1 SAR · Sentinel-2 · Tiling · Cloud masking · Temporal modelling · GeoPandas · Rasterio
- Edge
- Jetson AGX Xavier · Low-power CV · On-device perception · ARKit / LiDAR
07
Contact
Open to selective consulting and engineering work in geospatial machine learning, model optimisation and GPU infrastructure.