Deep learning engineer

Nadia Hamlaoui

Large-scale computer vision, generative AI, language models, and multimodal systems — taken from the lab notebook to GPU inference.

Selected work Get in touch
  • English C1
  • French C1
  • Arabic native
  • German A2

01 — Profile

From a facial signal to a system that ships.

I specialize in large-scale computer vision, generative AI, large language models, and multimodal systems. The work is high-performance machine learning: architectures, distributed training, and GPU inference, with a standing bias toward efficiency, scale, and the full path from research to production.

From 2022 to 2026 I worked on neural avatars — how a person is seen, reconstructed, lit, and rendered — and on the models and infrastructure around that problem.

  • 2022–2026 Machine learning engineer, Copresence GmbH
  • M.Eng. Information & communication technology, ENIT
  • MP* Mathematics & physics, Lycée Louis-le-Grand, Paris
  • DAAD Scholarship, TU Dresden — neuromorphic vision

02 — Work

Copresence GmbH

Machine Learning Engineer — Vision, Generative AI & ML Systems 2022–2026

01

Neural avatars, consistent from every view

Multi-view texture synthesis and delighting. DeepSets aggregate features without caring about view order. UV-space representations, multi-view projection, generative texture prediction, and color correction keep appearance stable as the camera moves.

Neural rendering tuned for multi-view appearance and real-time display. Studio light rebuilt in Blender with the Principled BSDF, then learned remapping matrices to correct subsurface-scattering and specular bias.

  • DeepSets
  • UV space
  • Neural rendering
  • Principled BSDF

02

Diffusion, inpainting, identity

Image generation and inpainting with latent diffusion, cross-attention conditioning, and transformers. Identity-preserving reconstruction trained with perceptual and adversarial losses, and PCA-based priors.

Facial and hair attributes modeled as distributions: uncertainty-aware objectives and directional statistics, including a von Mises loss, for estimates that stay reliable when the input is noisy or ambiguous.

  • Latent diffusion
  • Transformers
  • von Mises

03

Faces, read in real time

Multi-task facial analysis — segmentation, surface normals, canonical geometry, landmarks, and attributes — on shared representations. A YOLO-inspired FaceBoxes detector, trained with domain randomization, holds up under pose, scale, occlusion, and light.

ResNets pretrained on domain facial data outperformed ImageNet initialization after fine-tuning. Skin and eye color via self-training and pseudo-labels, with an annotation tool and a curated dataset behind the supervision.

  • Multi-task vision
  • FaceBoxes
  • Self-training

04

Geometry, measured against the studio

A validation system for 3D avatars: studio-captured ground truth, multi-view depth fusion, signed surface distances, collision detection, UV-space error heatmaps, region-level statistics, and automated multi-view visualization.

  • Depth fusion
  • Surface error
  • Evaluation

05

Vision, language, and speech in one stack

Multimodal systems joining vision models, large language models, retrieval-augmented generation, speech, and neural rendering. LLMs adapted with LoRA and PEFT, with retrieved context injected at inference.

  • RAG
  • LoRA / PEFT
  • Speech

06

From the cluster to the frame

Large-scale training on PyTorch Distributed Data Parallel: mixed precision, gradient checkpointing, and multi-GPU runs. Inference taken down in latency with TensorRT, ONNX Runtime, TorchScript, CUDA, operator fusion, quantization, and attention KV-cache optimizations.

  • PyTorch DDP
  • FP16 / AMP
  • TensorRT
  • INT8
  • KV-cache

03 — Research

Selected academic projects

TU Dresden, DAAD scholarship

Event-based gesture recognition

Neuromorphic vision models using spiking neural networks and event-based processing. Low-power edge AI for gesture recognition.

Acredius

AI cybersecurity systems

Machine-learning intrusion detection, combining supervised and unsupervised methods to surface malicious activity in network traffic.

04 — Talks

Conferences

Conference appearances, in citation form.

“A strong focus on model efficiency, scalability, and end-to-end ML infrastructure.”

Nadia Hamlaoui, profile
  1. 01

    Hamlaoui, Nadia.

    Mobile World Congress (MWC).

    Barcelona, Spain. Talk.

  2. 02

    Hamlaoui, Nadia.

    3DBody.Tech.

    Lugano, Switzerland. Talk.

  3. 03

    Hamlaoui, Nadia.

    German Conference on Pattern Recognition (GCPR).

    Heidelberg, Germany. Talk.

  4. 04

    Hamlaoui, Nadia.

    Cannes FilmTech.

    Cannes, France. Talk.

05 — Path

Training, tools, and the record.

Machine learning & generative AI

Computer vision, multimodal AI, vision-language models, diffusion models, GANs, latent diffusion, representation learning, neural rendering, self-supervised learning, probabilistic modeling.

LLMs & transformers

Transformers, vision transformers, RAG, embedding models, cross-attention, LoRA and PEFT, KV-cache optimization, token streaming.

Training

PyTorch DDP, mixed precision (FP16 / AMP), gradient checkpointing, multi-GPU training, CUDA optimization.

Inference

TensorRT, ONNX Runtime, TorchScript, CUDA profiling, quantization (FP16 / INT8), operator fusion, low-latency inference.

Libraries

PyTorch, TensorFlow, OpenCV, scikit-learn, FAISS, Hugging Face Transformers, Diffusers.

Programming

Python, C++, CUDA, Linux, Git, Shell, YAML.

Education

  1. 2019 — 2022

    M.Eng. in Information & Communication Technology

    National Engineering School of Tunis (ENIT), Tunisia

  2. Preparatory classes

    Mathematics & Physics, MPSI / MP*

    Lycée Louis-le-Grand, Paris, France

Awards

  • DAAD Scholarship TU Dresden — neuromorphic vision research
  • Excellence Scholarship Top baccalaureate ranking, for studies abroad
  • US Young Physicists’ Tournament 5th place

Languages

  • Arabic Native
  • French C1
  • English C1
  • German A2

06 — Contact

Write to me.

I am interested in international roles in computer vision, generative AI, and machine learning systems.

This opens your mail app, addressed to Nadia, with your note filled in.