Deep Learning · Signal Processing · NLP & LLMs · Embedded AI
I build machine learning systems for images, signals and language, from data preparation and model training to evaluation, optimization and deployment.
My skills cover computer vision, signal and image processing, NLP and LLM fine-tuning, with Python, C++ and PyTorch. I combine mathematical modeling and numerical optimization with practical software engineering to develop reproducible ML pipelines and efficient inference applications.
I work across model development and embedded deployment, including ONNX inference, ROS 2 perception pipelines and hardware implementation with FPGA and RISC-V.
Open to engineering opportunities in France in applied machine learning, computer vision, signal processing, NLP and embedded AI.
| Area | Skills & technologies |
|---|---|
| Machine learning & deep learning | Supervised and unsupervised learning, neural networks, model training and evaluation, PyTorch, scikit-learn |
| Computer vision | Image classification, object detection, anomaly detection, image preprocessing, OpenCV |
| Signal & image processing | Spectral analysis, hyperspectral unmixing, matrix factorization, constrained optimization, NumPy, SciPy |
| NLP & LLMs | Sequence modeling, RNNs, GRUs, Transformers, Hugging Face, language model fine-tuning, SFT, DPO |
| Inference & deployment | Python, C++, ONNX Runtime, ROS 2, GStreamer, Docker, Linux |
| Embedded systems & hardware | C, Raspberry Pi, FPGA, VHDL, RISC-V |
Detect and localize anomalies in industrial images using normal training examples. Compare a convolutional autoencoder with a compact PatchCore variant on three MVTec AD categories.
- Fitted models, a local inspection app and per-image predictions.
- Evaluation of image/pixel AUROC, defect recall, false alarms and CPU latency.
- Recorded experiment settings, data checksums and tests.
View the project · Read the experiment
Estimate material spectra and their fractions in hyperspectral images using SPA + FCLS, Simplex NMF and Spatial NMF.
- 112 measured fits across real benchmark scenes and controlled synthetic mixtures.
- Compare spectral angles, abundance errors and reconstruction residuals.
- A local pixel inspector, an executed notebook and saved numerical results.
View the project · Read the experiment
Connect PyTorch training with handwritten C++ inference for MLP and CNN digit classifiers targeting Raspberry Pi.
- Implement neural network forward passes without a deep learning runtime.
- Integrate OpenCV preprocessing with a GStreamer video pipeline.
- Compare model behavior through training notebooks and a project report.