Seeking a 2027 end-of-studies internship abroad

I build AI that
sees, reasons,
and runs fast.

ICT engineering student at SUP'COM, working where computer vision, autonomous agents, and high-performance systems meet.

Portrait of Mohamed Mdhaffar
vision_pipeline.py
01  stream = RTSP(source)
02  frames = model.track(stream)
03  insights = agent.reason(frames)
04  deploy(edge=True)
● LIVE32.8 FPS
CUDA acceleration9.5×peak speedup on A100
Tunis, Tunisia ↗ worldwide

Selected work / 2025–26

Systems with a point
of view—and a purpose.

01 / COMPUTATIONAL IMAGING · GPU SYSTEMS

GPU-Accelerated Deep Image Prior for Image Denoising

Single-image reconstruction with Deep Image Prior, ADMM, TV/WTV regularization, and native PyTorch C++/CUDA operators—built to study both reconstruction behavior and GPU execution.

33.049 dBBest ADMM-DIP checkpoint
9.52×Periodic-gradient speedup
A100 80 GBProfiling platform
Research repositoryExplore the implementation
Deep Image Prior optimization diagram showing a fixed noise input progressively reconstructed toward a target image
A randomly initialized network learns the structure of one image directly—without a pre-trained denoising dataset.

01 · Problem & motivation

Can the network architecture itself act as an image prior?

Deep Image Prior (DIP) fits an untrained convolutional network to a single noisy observation. Natural image structure is typically reconstructed before the network begins fitting noise, making optimization control and early stopping essential.

02 · Technical approach

DIP meets constrained optimization

I implemented vanilla DIP, then coupled it with ADMM—an optimization method that separates the neural reconstruction step from explicit TV/WTV regularization. Native CUDA operators cover periodic gradient, divergence, shrinkage, and dual updates, with autograd integration and a portable PyTorch fallback. I compared eager PyTorch, compiled PyTorch, and native CUDA paths, evaluated BM3D and other classical baselines, and added early stopping with best-checkpoint restoration to limit noise overfitting.

03 · Representative results

One RGB example from BSDS300

Each figure reads left to right: reconstructed output, noisy observation, and ground truth. These are representative single-image results—not dataset-wide averages or state-of-the-art claims.

BM3D result on one BSDS300 image: reconstruction on the left, noisy observation in the center, and ground truth on the right
BM3D baseline31.721 dB PSNR · 0.8032 SSIM
Vanilla Deep Image Prior result at iteration 723: reconstruction on the left, noisy observation in the center, and ground truth on the right
Vanilla DIP · iteration 72329.860 dB PSNR · checkpoint restored from iteration 703
ADMM-DIP result at iteration 1965: reconstruction on the left, noisy observation in the center, and ground truth on the right; best checkpoint reached 33.049 dB
ADMM-DIP · best checkpoint33.049 dB at iteration 1865 · displayed iteration 1965: 31.642 dB
CUDA-accelerated ADMM-DIP result at iteration 1999: reconstruction on the left, noisy observation in the center, and ground truth on the right
CUDA ADMM-DIP · iteration 199932.610 dB PSNR · CUDA accelerates computation, not reconstruction quality
Representative metrics for one BSDS300 image
MethodIterationPSNRgtPSNRnoisySSIMgtSSIMnoisyDSSIM
BM3D—31.721 dB27.4850.80320.54480.2778
DIP72329.860 dB26.6200.7607—0.2178
ADMM-DIP196531.642 dB best: 33.04927.513———
CUDA ADMM-DIP199932.610 dB27.913———

04 · CUDA acceleration

A 9.52× faster periodic-gradient operator

The benchmark compares the custom CUDA operator with the FFT implementation on a real DIP output of shape 1 × 3 × 320 × 480, using float32 on an NVIDIA A100. The result measures 100 operator calls.

FFT gradient391.895 µs39.189503 ms total
Custom CUDA41.165 µs4.116480 ms total

70.728 → 8.105 msProfiled self-CUDA time

Product image representing an NVIDIA A100 Tensor Core GPU
NVIDIA A100 80 GBAmpere architecture · 80 GB HBM2e · TF32, FP16 & BF16 Tensor Cores · high-bandwidth memory for optimization and profilingPlatform illustration; not asserted to be the exact physical machine used.

05 · Numerical consistency

Speed without changing the operator

The custom kernel stays closely aligned with the reference FFT computation. CUDA’s demonstrated contribution here is computational acceleration, not improved denoising quality.

Horizontal max error3.2619573e−07
Horizontal mean error6.2645512e−08
Vertical max error2.5774352e−07
Vertical mean error4.7060624e−08

06 · Profiling & performance analysis

From convergence curves to kernel-level traces

PyTorch Profiler and Perfetto made bottlenecks observable across Python, autograd, and CUDA rather than relying on wall-clock timing alone.

Loss and ground-truth PSNR curves comparing TV and WTV with eager PyTorch and custom CUDA implementations
Optimization behaviorLoss and ground-truth PSNR curves for TV/WTV eager and custom CUDA paths.
Perfetto execution timeline showing DIP forward passes, backward computation, optimizer steps, ADMM iterations, and CPU and CUDA activity
Perfetto execution traceThe trace exposes DIP forward passes, backpropagation, optimizer steps, ADMM iterations, and CPU/CUDA activity.

07 · Technologies

Built across optimization and systems layers

  • Python
  • PyTorch
  • C++
  • CUDA
  • ADMM
  • Deep Image Prior
  • TV / WTV
  • BM3D
  • PyTorch Profiler
  • Perfetto
  • Jupyter

08 · Limitations & future work

What comes next

The reported metrics come from one RGB BSDS300 example and do not establish dataset-wide performance. Chest-CT reconstruction, a CT dataset, and a CT measurement operator remain future work despite the repository title. Next steps include broader evaluation, task-specific measurement models, and systematic quality/runtime studies.

02 / SPACE SYSTEMS · EDGE AI
2nd place · TSYP13

AI-Driven Predictive Maintenance for OPS-SAT ADCS

An autonomous monitoring prototype for a 3U CubeSat Attitude Determination and Control System. It detects abnormal telemetry and estimates corrected sensor values directly on embedded hardware, reducing dependence on ground intervention.

95.65%ANN test accuracy
89.10%F1 score
97.63%Test AUC
≈2 msEmbedded ANN latency
Project repositoryExplore the models, firmware, and schematics
OPS-SAT spacecraft illustration beside a labeled view of its internal bus system stack
OPS-SAT mission context and compact 3U bus architecture.

01 · Mission challenge

Keeping attitude-control telemetry dependable in orbit

CubeSat ADCS hardware operates under strict power, volume, thermal, and radiation constraints. A sensor anomaly can threaten pointing accuracy and mission continuity, while waiting for ground analysis adds delay. This project investigates onboard intelligence for faster fault detection and autonomous signal recovery.

02 · Dual-model strategy

Detect first, then reconstruct the signal

An Artificial Neural Network classifies incoming telemetry as normal or anomalous. When a fault is detected, one of nine channel-specific Random Forest regressors estimates a corrected value for the affected magnetometer or sun-sensor channel.

03 · End-to-end architecture

Telemetry → detection → correction → validated output

The STM32F429 sits inside the monitoring loop, consuming ADCS signals, running anomaly inference, and forwarding corrected data back toward the simulated bus interface.

Architecture diagram showing STM32F429 telemetry acquisition, neural-network anomaly detection, Random Forest correction, and communication with OPS-SAT ADCS components
The prototype connects the STM32F429, NanoCom AX100 simulation, magnetometer channels, six coarse sun sensors, and the spacecraft attitude-control bus.
3Magnetometer channelsX, Y, and Z
6Sun-sensor channelsPhotodiodes PD1–PD6
9Correction modelsOne Random Forest per channel

04 · Machine learning

Compact models for constrained inference

The pipeline separates fault classification from numerical signal correction, allowing each model to remain focused and small enough for a Cortex-M4F deployment target.

MODEL 01 · DETECTION
Artificial Neural Network

Analyzes real-time magnetometer and sun-sensor telemetry to identify deviations from normal behavior.

  • 95.65% accuracy
  • 89.10% F1
  • 97.63% AUC
MODELS 02–10 · CORRECTION
Nine Random Forest regressors

Each regressor specializes in one monitored channel and predicts a corrected signal when the detector flags an anomaly.

  • Per-channel models
  • Autonomous recovery
  • No immediate ground intervention

05 · Embedded deployment

Inference on STM32F429

The final firmware targets the STM32F429ZIT6 Discovery Kit with an Arm Cortex-M4F. STM32Cube.AI converts the ANN into optimized C for embedded inference, while Edge Impulse supports rapid experimentation and comparison.

≈2 msANN inference latency
≈0.2 KBRAM usage
≈34.7 KBFlash usage
Physical CubeSat-style prototype frame containing an STM32F4 Discovery board and connected sensor wiring
Physical embedded prototypeSTM32F4 Discovery hardware integrated inside the CubeSat-style test frame.
MCUSTM32F429ZIT6Onboard processing and inference
Attitude sensingMPU-9250 IMU9-axis inertial measurement
CommunicationsNanoCom AX100Subsystem interface simulation

06 · TSYP13 outcome

The project combined machine learning, embedded deployment, telemetry processing, and hardware integration into one demonstrable prototype for the TSYP13 challenge.

07 · Technologies

ML, embedded, and electronics

  • TensorFlow
  • Keras
  • scikit-learn
  • STM32F429
  • Arm Cortex-M4F
  • STM32Cube.AI
  • STM32CubeIDE
  • Edge Impulse
  • Embedded C
  • ADCS telemetry

08 · Engineering scope

A mission-oriented prototype

The work demonstrates the detection, correction, and embedded deployment pipeline for OPS-SAT-inspired ADCS monitoring. It is presented as a prototype and competition project—not as software already flown aboard the ESA spacecraft.

03 / BIOMEDICAL DISCOVERY · VECTOR SEARCH
7th place · Vectors in Orbit Hackathon

Bio-Vector Orbit Discovery Engine v2.0

A vector-powered discovery engine that turns fragmented biomedical sources—research papers, protein structures, and chemical molecules—into one searchable intelligence layer.

LiteratureSemantic PubMed retrieval
ProteinsPDB structure discovery
MoleculesSMILES latent similarity
Project repositoryExplore Bio-Vector Orbit v2.0
Bio-Vector Orbit dashboard showing semantic protein search results for tyrosine with cards linking to interactive 3D views
Protein discovery mode combines semantic matches, structural metadata, and direct access to interactive 3D visualization.

01 · Problem

Biological knowledge lives in disconnected formats

Researchers move between article abstracts, sequence and protein repositories, and chemical representations. Different identifiers, data shapes, and search interfaces make it difficult to explore related evidence through one consistent workflow.

02 · Solution

A shared semantic layer across biological entities

Bio-Vector Orbit standardizes retrieval behind one interface. Users choose a discovery mode, submit text or a SMILES representation, and receive specialized results for literature, protein structures, or chemically similar molecules.

03 · Discovery modes

Three search experiences, one interface

The backend normalizes each result type while the React interface adapts its cards and actions to the biological object being explored.

01
Biomedical literature

Retrieves PubMed records through NCBI and embeds biomedical text with BGE-Small for meaning-based discovery beyond exact keywords.

NCBI · PubMed · Biopython · FastEmbed
02
Protein structures

Searches RCSB PDB structures and opens selected proteins inside a PDBe Mol* viewport with structural metadata and an immersive HUD.

RCSB PDB · PDBe Mol* · 3D visualization
03
Chemical molecules

Encodes SMILES strings through a custom Keras variational autoencoder and searches the resulting latent vectors for molecular similarity.

SMILES · Keras VAE · PubChem

04 · System architecture

From query to specialized scientific result

A FastAPI service coordinates external biological sources, embedding models, and Qdrant collections, while Docker Compose starts and connects the application stack.

Literature sourceNCBI PubMedReal-time article retrieval
Structure sourceRCSB PDBProtein records and coordinates
Chemical sourcePubChem + VAESMILES-based latent discovery

05 · Vector intelligence

Purpose-built representations for each modality

BGE-Small embeddings support semantic literature and protein queries. Chemical search follows a separate learned representation: the custom VAE maps SMILES inputs into a latent space where nearby vectors indicate similarity. Qdrant keeps these vector types in dedicated collections and applies HNSW indexing for efficient nearest-neighbor retrieval.

BGE-SmallText and protein discoveryFastEmbed inference
Custom VAEMolecular latent spaceKeras model
HNSWNearest-neighbor indexingQdrant search

06 · User workflow

A focused path from question to exploration

The same interaction model supports very different scientific data while preserving modality-specific context.

  1. 01
    Choose a mode

    Switch between Literature, Proteins, and Molecules.

  2. 02
    Run neural retrieval

    Embed a query or encode a SMILES string, then search the appropriate vector collection.

  3. 03
    Render the right object

    Receive adaptive PubMed, PDB, or molecule result cards.

  4. 04
    Explore in context

    Inspect metadata or open a protein in the interactive 3D viewer.

07 · Applications

Designed for scientific discovery workflows

  • Drug discoveryExplore literature, target proteins, and related compounds.
  • Literature reviewSearch biomedical abstracts by meaning rather than exact wording.
  • Hypothesis generationSurface relationships across biological concepts and structures.
  • EducationMove from a biological query to visual protein exploration.

08 · Technologies

A containerized multimodal stack

  • React 19
  • TypeScript
  • Tailwind CSS
  • FastAPI
  • Python 3.11
  • Qdrant
  • FastEmbed
  • BGE-Small
  • Keras VAE
  • PDBe Mol*
  • Docker Compose
  • NCBI
  • RCSB PDB
  • PubChem

09 · Project scope

Hackathon-ready discovery engine

The project demonstrates a unified interface and retrieval architecture across three biological modalities. Its value is exploratory search and visualization; it is not presented as a clinical, diagnostic, or validated drug-development system.

04 / FINAL PROJECT · MEDICAL AI
Two-phase pipeline complete

Robust Multi-Class Diabetic Retinopathy Detection

A hybrid deep-learning pipeline that grades diabetic retinopathy from retinal fundus images, adds model-trust safeguards, and packages predictions with privacy-oriented encryption and watermarking controls.

93.0%Validation accuracy
0.9566Quadratic Weighted Kappa
5 classesDR severity grading
Project repositoryExplore the complete implementation
Medical illustration comparing a normal retina with diabetic retinopathy signs including haemorrhages, abnormal blood vessels, aneurysms, cotton-wool spots, and hard exudates
Diabetic retinopathy can produce visible retinal lesions that a computer-vision system can learn to distinguish by severity.

01 · Clinical problem

Grading a progressive retinal disease

Diabetic retinopathy screening requires distinguishing subtle lesions across five ordered severity levels: No DR, Mild, Moderate, Severe, and Proliferative DR. The problem combines fine local texture recognition with global retinal context and substantial class imbalance.

02 · Design goal

Accuracy, ordinal agreement, and responsible outputs

The project combines a hybrid CNN–Transformer classifier with explainability, uncertainty estimation, and out-of-distribution checks. A second security phase protects patient metadata and predictions inside a DICOM-oriented workflow.

03 · End-to-end architecture

From fundus image to interpretable severity prediction

The 512 × 512 retinal input passes through image preparation and augmentation before a ConvNeXt–Swin hybrid combines local lesion features with global spatial relationships.

Diabetic retinopathy pipeline diagram showing retinal datasets, preprocessing and augmentation, hybrid CNN and Vision Transformer feature extraction, five-class output, and Grad-CAM explainability
Project pipeline overview: data preparation, hybrid CNN–ViT inference, five-class grading, and visual explanation.

04 · Training pipeline

Built for retinal variation and class imbalance

The model has approximately 112.8 million trainable parameters. Training combines robust retinal preprocessing, strong augmentation, imbalance-aware loss, and ordinal evaluation.

PREPROCESSING
Preserve useful retinal detail

ROI crop, Ben Graham normalization, green-channel CLAHE, and resizing to 512 × 512.

AUGMENTATION
Increase visual diversity

Full-range rotation, horizontal and vertical flips, subtle zoom, MixUp α=0.4, and CutMix α=1.0.

OBJECTIVE
Handle difficult minority classes

Focal Loss γ=2 with label smoothing ε=0.1 and inverse-frequency class weights.

OPTIMIZATION
Stable end-to-end training

AdamW with a 1e−4 learning rate, 1e−5 weight decay, and cosine-annealing scheduling.

05 · Validation results

Best checkpoint at epoch 26 of 50

QWK is the primary metric because it rewards agreement that respects the ordered distance between DR severity classes. These values are validation results for the project’s 6,399-image dataset—not evidence of clinical efficacy.

Validation accuracy93.0%
Quadratic Weighted Kappa0.9566
Training loss0.3081
Validation loss0.2617
Learning rate4.69 × 10⁻⁵
Time per epoch≈285 s
Class distribution · 6,399 fundus images
Severity classImagesShare
No DR2,80543.8%
Mild74011.6%
Moderate1,89829.7%
Severe3665.7%
Proliferative DR5909.2%

06 · Model trust

Safeguards around the prediction

A high validation score alone is not enough for responsible medical AI. The inference layer adds visual evidence, confidence analysis, and input-validity checks so uncertain or unfamiliar cases can be surfaced rather than silently treated as reliable.

Grad-CAM++Highlights retinal regions influencing the decision.
Attention rolloutSummarizes Transformer attention across the image.
MC DropoutEstimates predictive uncertainty over repeated passes.
OoD detectionFlags inputs outside the expected retinal-image distribution.

07 · Medical data security

Encrypted metadata and invisible watermarking

Phase 2 is a privacy-oriented engineering prototype. It implements concrete security mechanisms but is not described as independently audited HIPAA compliance.

01ROI extractionSeparate the retinal region from the black background.
02AES-256 encryptionEncrypt patient identifiers and prediction payloads.
03DWT watermarkingEmbed encrypted data within the region of non-interest.
04DICOM packagingExport the image, metadata, and integrity information.
SHA-256Integrity verification
>40 dBReported watermarked-image PSNR
Reverse extractionRecover embedded data and optional JSON reporting

08 · Technologies

Vision, trust, and security stack

  • PyTorch
  • ConvNeXt
  • Swin Transformer
  • OpenCV
  • timm
  • scikit-learn
  • Grad-CAM++
  • PyWavelets
  • AES-256
  • SHA-256
  • PyDICOM
  • Jupyter

09 · Scope & limitations

A completed research implementation—not a clinical product

The two software phases are implemented and validated on the project dataset, but the reported results do not establish real-world clinical safety, regulatory approval, or HIPAA compliance. External validation, calibration studies, bias analysis, security auditing, and prospective clinical evaluation would be required before deployment.

Implementation and paper by Khalifa Bouneb & Mohamed Mdhaffer.

See experiments and code on GitHub ↗

Experience

From a research idea
to a running system.

About / Mohamed Mdhaffar

Engineering intelligence
for the real world.

I’m a final-year ICT engineering student at SUP’COM in Tunis. I care about the whole AI stack: the model, the system around it, and the hardware it ultimately runs on.

My current interests include efficient LLM inference, GPU optimization, computer vision, AI-agents and neurosymbolic testing of deep-learning libraries.

Education

A rigorous foundation,
built to go further.

Higher School of Communications of Tunis — SUP’COM campus
2024—2027 · TUNIS

ICT Engineering

Higher School of Communications of Tunis — SUP’COM

At SUP’COM, I am developing a broad engineering foundation across telecommunications, computer networks, software systems, data, and artificial intelligence. My studies connect theory with practical implementation through technical projects, teamwork, and the design of intelligent systems that can operate under real-world constraints.
2022—2024 · SFAX

Mathematics & Physics Branch

Preparatory Institute for Engineering Studies of Sfax

This demanding two-year program gave me a rigorous scientific foundation in mathematical analysis, linear algebra, statistics, probability, numerical methods, mechanics, electromagnetism, thermodynamics, chemistry, computer science, and engineering fundamentals. The intensive pace strengthened my discipline, analytical thinking, resilience, and ability to perform under pressure.I ranked 74th out of 1,750 candidates (top 5%) in Tunisia’s highly selective national engineering entrance exam, earning admission to SUP’COM.

Communication

Languages

ع

Arabic

Mother tongue

Native
Fr

French

Fluent communication

Bilingual
En

English

TOEIC · 765

Professional

Community

Volunteering & leadership

Farah Association for the Integration of Children with Autism and Special Needs logo
Farah AssociationIntegration of children with autism and special needs
TECH FOR GOOD · DEVELOPMENT TEAM

AutiSens

A volunteer-driven initiative for autism support

Working with the Farah Association for the Integration of Children with Autism and Special Needs, our team built a mobile application that digitizes the PEP-3 psychoeducational assessment used during consultations. The platform helps therapists reduce paperwork, save time, and make evaluations more accessible and efficient for caregivers and families.

Beyond development, we created a public street-interview awareness campaign to challenge stigma, share reliable information about autism, and promote inclusion.

  • Flutter
  • Firebase
  • UX/UI
  • Social impact
Built with Iyed Rachdi, Tasnim Selmi, Malek Ben Jomaa, and Mohamed Aziz Saoudi.
SUP’COM IEEE Student Branch logo
SUP’COM IEEE Student BranchIEEE GRSS AESS JOINT Chapter
IEEE GRSS · HR MANAGER & INSTRUCTOR

Leadership through community

IEEE GRSS AESS JOINT Chapter - Sup'Com Student Branch

As HR Manager, I supported member engagement, team coordination, onboarding, and a collaborative culture across the student branch. The role strengthened my leadership, communication, organization, and ability to help multidisciplinary teams work effectively together.

I also designed and delivered a technical workshop on deep learning for supernova detection, translating complex AI concepts into an accessible hands-on learning experience for students.

  • Leadership
  • Team coordination
  • Deep learning
  • Technical instruction

Technologies & Skills

My technical
constellation.

Programming Languages

Python
C
C++
JavaScript
MATLAB

AI & Machine Learning

PyTorch
TensorFlow
scikit-learn
pandas
OpenCV

GPU & Edge AI

CUDACUDA C++
TritonTriton
NumbaNumba

Software, Backend & Cloud

HTML
CSS
Django
React
QdrantQdrant
Git
GitHub
Docker
AWS EC2
LangChainLangChain
LangGraphLangGraph
Model Context ProtocolMCP
LangSmithLangSmith

IoT & Embedded

Arduino
STMSTM32
ESP32
Raspberry Pi 4

Stanford course-based learning

Theory studied.
Ideas implemented.

Self-directed study using Stanford’s publicly available course materials, complemented by practical GPU programming work. Each course strengthens a different layer of my AI and high-performance computing foundation.

S
CS229 · STANFORD

Machine Learning

Supervised and unsupervised learning, linear and logistic regression, generalized linear models, SVMs and kernels, neural networks, clustering, dimensionality reduction, learning theory, reinforcement learning, and practical model evaluation.

  • Statistical learning
  • Optimization
  • RL & control
Official course ↗
S
CS230 · STANFORD

Deep Learning

Neural-network foundations, initialization, regularization, optimization, error analysis, convolutional architectures, sequence models, deep-learning project strategy, and the practical process of building reliable learning systems.

  • Neural networks
  • Optimization
  • ML strategy
Official course ↗
S
CS231N · STANFORD

Deep Learning for Computer Vision

Image classification, backpropagation, convolutional networks, training and fine-tuning, detection, segmentation, visual recognition architectures, generative approaches, and practical engineering for real-world vision systems.

  • CNNs
  • Detection
  • Visual recognition
Official course ↗
S
CME295 · STANFORD

Transformers & Large Language Models

Tokenization, embeddings, attention, Transformer architecture, positional encoding, BERT-style models, mixture of experts, prompting, sampling, context handling, LLM training, evaluation, and efficient adaptation techniques.

  • Attention
  • LLM training
  • Inference
Official course ↗
S
CS236 · STANFORD

Deep Generative Models

Probabilistic foundations of generative learning, autoregressive models, variational autoencoders, normalizing flows, generative adversarial networks, energy-based models, diffusion models, evaluation, and representation learning.

  • VAEs
  • GANs
  • Diffusion
Course materials ↗
S
CS149 · STANFORD

Parallel Computing

Multicore processors, SIMD execution, GPU programming, memory locality, synchronization, scheduling, parallel programming abstractions, performance analysis, and designing scalable implementations for modern hardware.

  • SIMD
  • GPU systems
  • Performance
Official course ↗
S
EE367 / CS448I · STANFORD

Computational Imaging

Digital photography, image signal processing, denoising, deconvolution, inverse problems, proximal optimization, CNN-based imaging, wave optics, time-of-flight sensing, and joint optimization of optics and algorithms.

  • Inverse problems
  • Image processing
  • Computational optics
Official course ↗
GPU Mode
ADDITIONAL PRACTICAL TRAINING

GPU Mode · CUDA

Hands-on GPU programming focused on CUDA kernels, memory hierarchy, profiling, performance bottlenecks, parallel primitives, kernel fusion, and systematic optimization of machine-learning workloads.

  • CUDA kernels
  • Profiling
  • Optimization
Cover of Programming Massively Parallel Processors: A Hands-on Approach by David B. Kirk and Wen-mei W. Hwu
Currently readingProgramming Massively Parallel ProcessorsDavid B. Kirk & Wen-mei W. Hwu
Independent practical work

Have a challenge worth solving?

Let’s build something
that matters.

mohamed.mdhaffar@supcom.tn ↗
Based in Tunis · Open to relocation
LinkedIn ↗GitHub ↗