Ali Aouf
CV

Research,
made useful.

Machine learning engineer & researcher.

Meet Ali
Cross-hatched portrait of Ali thinking with a notebook on an open Andalusian terrace overlooking the countryside
An olive tree overlooking a valley, drawn in sepia ink and watercolorRooted in place.
Always growing.

A researcher’s curiosity.
A builder’s instinct.

I’m Ali, a machine learning engineer and researcher. I like the space between understanding a difficult problem and building something useful from it.

My work has taken me from medical imaging and drilling operations in Algeria to deep generative models in Belgium. Across those settings, I’ve owned the whole path: research design, model development, production engineering, and the interfaces people use. I’ve passed my private PhD defence at UCLouvain, carried out at SCK CEN; the public defence is on 21 January 2027.

Based in
Mol, Belgium
Drawn to
Generative models & useful ML systems
Ali Aouf
5 yearsIn ML research & engineering
9 publicationsResearch across disciplines
Jan 2027PhD public defence · private defence passed
I.

Looking beneath the surface

From CT images to porous materials, I work with models that recover useful structure from incomplete observations.

II.

Staying with the problem

A model is one part of the work. Data pipelines, deployment, access control, and reliable interfaces matter too.

III.

Building with others

I’ve led a small engineering team, mentored interns, and connected industry projects with research laboratories.

One chapter
leads to
another.

From learning to see in medical images
to learning to generate in three dimensions.

0104
Algiers → Mol · 2021–present

Generating
what we can’t see.

Apr 2023 — Present · Mol, Belgium

Doctoral Researcher — Deep Generative Models
SCK CEN, in collaboration with UCLouvain

I build generative models that synthesise 3D microstructures from 2D cross-sections, reducing reliance on costly 3D tomography for porous-media characterisation.

The work, in detail
  • Designed DimExDAM, a diffusion-adversarial model that scores denoising transitions, and ControlledSliceFlow, a flow-matching model with a decomposed noise prior for cross-volume diversity and spatial porosity control. ControlledSliceFlow is open source.
  • Used two-stage reflow to reach reference-quality sampling in 8 function evaluations.
  • Built a Numba-accelerated D3Q19 lattice-Boltzmann solver: within 0.2% of the analytic square-duct solution and roughly 220× faster than a naive implementation.
  • Standardised a benchmark across five materials, including clays, a carbonate, and a sandstone.
  • Built reproducible training pipelines on a Slurm GPU cluster, with versioned configurations, unit tests and experiment tracking (Weights & Biases, Comet ML), plus a 3D visualisation tool for flow-matching ODE trajectories. All of it within SCK CEN’s nuclear-sector security and data-handling rules.
Private defence passed. Public defence: 21 January 2027.

From the first line
to the field.

Jul 2022 — Mar 2023 · Algiers, Algeria

Team Lead — Machine Learning
Smart Drilling Operations

I built and maintained SmartWells from scratch: a role-based platform connecting rig operations, reporting, costs, and operational insight.

The work, in detail
  • Delivered role-based access with an audit trail of user activity, daily activity and cost entry, an interactive well map, a well schematic viewer, a KPI dashboard, and seven report types. Designed for offline entry at the rig with automatic synchronisation.
  • Ran delivery on GitHub: pull requests with code review, GitHub Actions CI with pytest, and separate development, test and production environments. Near-daily releases at first, then weekly and monthly release trains.
  • Owned rate-of-penetration optimisation from feasibility study to model development, supervising interns on supporting components.
  • Supervised and reviewed Gazoil, a fuel-monitoring product covering tank levels, refill history, noisy consumption data, and forecasting.
  • Grew the team from two people to eight at peak, usually two engineers reporting to me plus interns. Led technical direction, code review and mentoring, and served as the main technical contact for client demonstrations, including the Social and Economic Forum in Algiers in February 2023.
Research, product ownership, and team leadership in the same chapter.

Models that meet
the world.

Oct 2021 — Feb 2023 · Algiers, Algeria

Machine Learning Engineer
Alpha Computers SPA, parent company of Smart Drilling Operations

I worked across industrial data, accessibility, and conversational systems—building the software around the models as well as the models themselves.

The work, in detail
  • Built real-time drilling-data ingestion and storage; standardised LAS/DLIS and legacy formats to WITSML 2.0 and streamed data through Energistics ETP.
  • Led a joint project with CERIST on a sensory-substitution navigation aid for visually impaired users, translating visual input into auditory and haptic modalities. The software worked; hardware availability remained the blocker.
  • Delivered conversational ML systems for insurance clients, including a Rasa-based claims and policy-advisory assistant supporting guided and open-ended dialogue.
A bridge between a research lab, an engineering team, and its users.

Learning to see.
Learning to simplify.

Feb — Oct 2021 · Algiers, Algeria

Research Intern — Medical Imaging
CDTA · Centre for Development of Advanced Technologies

I designed O-Net for automatic segmentation of COVID-19 lesions in chest CT, pursuing an architecture that was smaller and faster to train.

The work, in detail
  • Reworked U-Net with a second upsampling channel while removing costly inter-channel skip connections.
  • Reached 476K parameters against U-Net’s 31M, with convergence in a quarter of the epochs and a Dice score of 0.866.
  • Implemented and evaluated the model across three public CT corpora, with better cross-scanner generalisation. Work published in Computers & Graphics and at IEEE ISPA.
A research question carried through implementation, evaluation, and publication.

Learning is
part of the work.

PhD in Artificial Intelligence

2023 — 2027

UCLouvain, in collaboration with SCK CEN · Belgium

Deep Generative Models for 3D Microstructure Synthesis

Private defence passed · Public defence 21 January 2027

MSc, Smart Computational Systems

2019 — 2021

USTHB · Algiers, Algeria

Graduated with honours

BSc, Computer Science

2016 — 2019

University of Constantine 2 Abdelhamid Mehri · Algeria

Top 5 of class

A good question
deserves a working answer.

Understand. Experiment. Build. Validate.

Three questions.
Three kinds of impact.

From the structure of materials
to the systems people rely on.

From 2D
to 3D.

The structure
we cannot see.

Doctoral research · DimExDAM & ControlledSliceFlow · 2023–present

How do you reconstruct a material in three dimensions when you only have two-dimensional cross-sections? I explore that question with diffusion, adversarial learning, and flow matching.

My role
Research design & implementation
Context
SCK CEN × UCLouvain
8 function evaluations for sampling5 materials in a unified benchmark
Follow the research

Beyond a convincing image

The work spans controllable spatial porosity, diversity across generated volumes, and the evaluation of physical transport properties—not just visual similarity.

Building the test as well as the model

I wrote a Numba-accelerated D3Q19 lattice-Boltzmann solver for permeability and diffusivity benchmarking. It was validated to within 0.2% of the analytic square-duct solution and ran roughly 220× faster than a naive implementation.

Making experiments reproducible

GPU-cluster training, experiment tracking, a shared protocol across five materials, and a 3D tool for inspecting flow-matching ODE trajectories support the research.

Read the 2026 paper ↗ Explore the ControlledSliceFlow code ↗
Built for
the field.

SmartWells.
The whole system.

Smart Drilling Operations · 2022–2023

A rig does not stop working when its connection drops. I built SmartWells to bring drilling, workover, and snubbing operations into one platform, with offline entry and automatic sync.

My role
Builder, maintainer & team lead
Scope
Full-stack operational platform
7 report types generatedOffline entry, then automatic sync
Explore the build

One place for the daily operation

The platform combined role-based access, activity and cost entry, an interactive well map, well schematics, and KPI dashboards.

A product that kept growing

I supervised Gazoil, a fuel-monitoring product extended from the SmartWells codebase, covering tank levels, refill history, consumption analysis over noisy sensor data, and forecasting.

Shipped as a team

Pull requests with code review, GitHub Actions CI with pytest, and separate development, test and production environments. Releases went from near-daily patches to weekly and monthly trains as the product matured, and the team grew from two people to eight at peak.

More than implementation

I led technical direction, reviewed code, mentored interns, and represented the team in client demonstrations. I also owned rate-of-penetration optimisation from feasibility to model development.

Cross-hatched stone arcade with a narrow shaft of sunlight illuminating a single stone
Less weight.
More focus.

O-Net.
Rethinking the model.

CDTA · Medical imaging · 2021

For COVID-19 lesion segmentation in chest CT, I explored how an architectural change could make a model much smaller without losing the detail that mattered.

My role
Architecture, implementation & evaluation
Evaluation
Three public CT corpora
476K parameters vs U-Net’s 31M0.866 Dice score
Look inside the research

A different route through U-Net

O-Net introduces a second upsampling channel while removing costly inter-channel skip connections.

Efficiency with evidence

The model converged in a quarter of the epochs compared with U-Net and showed better cross-scanner generalisation across the evaluated CT corpora.

A shared research contribution

The work was published with collaborators in Computers & Graphics and at IEEE ISPA. The architecture and evaluation grew out of my research internship at CDTA.

Read the COVIR paper ↗

Ideas I keep
coming back to.

Personal projects in language, information, and data-sovereign AI.

NewsForge / MOVET

Multilingual information

A news-to-narrative platform ingesting 405 sources across 37 languages, connecting reporting through an Article → Claim → Event → Thread ontology.

Inside the idea

The LLM pipeline handles cross-document event coreference and cross-lingual merging, bringing related reporting into a shared narrative structure.

Document OCR

Document intelligence

Entity extraction from phone-camera photographs of official identity documents, including mixed Arabic and Latin script.

Inside the system

An image-quality gate, per-document JSON schemas, and self-hosted vLLM inference support the workflow under a data-residency constraint.

SaharaAI

Data-sovereign deployment

An ML-as-a-service platform built around self-hosted inference and multi-tenant service design.

Inside the stack

FastAPI, Keycloak, PostgreSQL, and React form the foundation for application delivery and access control.

Questions worth
sharing.

My publications span generative modelling, materials science, medical imaging, and Arabic speech data.

9publications
across disciplines
Explore my Google Scholar profile ↗

Publication count from my September 2026 CV.

2025

3D clay microstructure synthesis using denoising diffusion probabilistic models.

Aouf A, Laloy E, Rogiers B, De Vleeschouwer C.

Applied Computing and Geosciences
2025

Low cycle fatigue life prediction for neutron-irradiated and non-irradiated RAFM steels via machine learning.

Zahran H, Zinovev A, Terentyev D, Aouf A, Wahab MA.

Fusion Engineering and Design
2022

Augmented reality for COVID-19 aid diagnosis: CT-scan segmentation based on deep learning.

Amara K, Aouf A, Kerdjidj O.

IEEE ISPA
2019

Basic Arabic Vocal Emotions Dataset (BAVED).

Aouf A.

Dataset
In conversation

Talks & posters at InterPore 2026, GeoStats 2024, and RFIAP 2024.

Research depth.
Engineering range.

I.

Generative AI

From model foundations to inference.

  • Diffusion · DDPM / DDIM
  • Flow matching · GANs · VAEs
  • Transformers & LLMs
  • Fine-tuning · RAG · Agentic workflows
  • vLLM · Self-hosted inference
II.

ML & computation

The tools behind the experiments.

  • PyTorch · TensorFlow · Keras
  • Hugging Face · scikit-learn · OpenCV
  • Python · Numba / CUDA
  • GPU-cluster training
  • Weights & Biases · Comet ML
III.

Production systems

The engineering around the model.

  • FastAPI · Docker · Model serving
  • Git · GitHub Actions CI · pytest
  • Keycloak / OIDC · Audit logging · Telemetry
  • PostgreSQL · MySQL · Redis · Spark
  • Go · Java · JavaScript / React

Follow the question.
Own the outcome.

Research design, implementation, and delivery belong in the same conversation.

01

Understand what matters.

Start with the problem, the available evidence, and the constraints. A feasibility study can be as important as the model that follows.

02

Make the idea testable.

Build clear experiments, useful baselines, and evaluation that means something—from cross-scanner generalisation to physical transport benchmarks.

03

Stay through the delivery.

Connect the model to data pipelines, deployment, and a usable product. Review the code, support the team, and keep the full system in view.

More than one
way to connect.

Arabic
Native
English
Fluent · IELTS
French
Fluent
Spanish
Intermediate
Dutch
Elementary · improving

The door
is open.

Have a research question, an ML challenge,
or a system worth building?
I’d be glad to talk.

Email me

Based in Mol, Belgium. Open to machine learning engineering and research roles.

LinkedIn ↗GitHub ↗Google Scholar ↗Download CV ↓