Sinda Besrour
Software Engineer
From scalable machine learning backends to edge-ready IoT integrations
Turning complex tech into real-world impact.
Previously C++ Lecturer at Université de Moncton and Software Developer at Talan
About Me
I'm a Software Engineer with a Master's in Computer Science, specializing in AI-driven solutions, cloud infrastructure, and IoT systems. My work bridges cutting-edge research and real-world deployment—whether it's building scalable ML pipelines with PyTorch and FastAPI, deploying federated learning models with Flower, or creating real-time apps using React and Unity.
With hands-on experience in AWS, Docker, Postman, and Nginx, I build backend systems that are fast, secure, and ready for scale. I'm passionate about bringing intelligence to infrastructure—connecting LLMs, LVMs, and edge devices into cohesive, automated workflows. Whether it's a cloud-based API or an embedded device, I believe good engineering should always feel seamless.
My Stack
AI, Deep Learning & Federated Systems
LLMs (GPT, BERT), Transformers, Federated Learning with Flower, Transfer Learning (ResNet, EfficientNet).
Tools & Libraries:
PyTorch, TensorFlow, Keras, Scikit-learn, Pandas, NumPy, Hugging Face, Flower
Backend & Cloud Engineering
Building and deploying ML APIs, real-time backend services, and distributed infrastructure on the cloud.
Tools & Libraries:
Flask, FastAPI, Node.js, Express.js, AWS (EC2, RDS), Azure AI, Nginx, Docker, PM2
Metaverse & Game Development
Creating multiplayer games and interactive experiences using Unity and Photon, with real-time communication powered by Agora.
Tools & Libraries:
Unity 3D, C# Photon PUN, Photon Realtime, Photon Chat Agora SDK (Voice, Video, Streaming)
Computer Vision & Signal Processing
Visual and audio processing for intelligent systems, LVM-based classification, and gesture/audio recognition.
Tools & Libraries:
OpenCV, Pillow, Torchvision, Librosa, SciPy, Torchaudio
Web & Mobile Interface Development
Designing modern, responsive, and accessible web and mobile interfaces with a focus on performance and interactivity.
Tools & Libraries:
React.js, React Native, HTML, CSS, JavaScript, Redux, Axios
API Integration & System Tools
CI/CD pipelines, API testing, version control, deploying across Linux and embedded platforms.
Tools & Libraries:
Postman, GitHub/GitLab, Linux (Ubuntu, Kali), MacOS, Raspberry Pi, ESP32, Arduino
My Projects
Mitigation of Poisoning Attack Coalitions in Federated Learning
Designed a custom non-linear federated learning aggregation algorithm. Aimed to mitigate the effect of poisoning attack coalitions on model performance.
Tools Used:
- Flower, PyTorch, NumPy, Scikit-learn, Pandas, Matplotlib, Seaborn
Key Achievements:
- Attack Coalition Impact:
- Accuracy dropped to 18.34%.
- F1-score dropped to 18.68%.
- Post-Mitigation Results:
- Accuracy improved to 99.36%.
- F1-score improved to 98.46%.
Research Focus & Innovation:
- Tackled ethical issues in data privacy and model integrity in decentralized AI.
- Promoted responsible AI by countering large-scale coordinated attacks.
Federated Network
Generalization vs Personalization: Mitigating Data Heterogeneity in Federated Learning
Designed a linear aggregation strategy that weighs local models by training loss and data quality to reduce the effect of sensor data heterogeneity, followed by model interpolation for client-side personalization.
Tools Used:
- Flower, Scikit-learn, Xgboost, Numpy, Pandas, Matplotlib, Seaborn
Key Achievements:
- Reached 88% average F1-score.
- Outperformed baselines with over 94.3% improvement.
- Handled both balanced and one-label clients effectively.
Research Focus & Innovation:
- Explored how FL systems can balance generalization and personalization.
- Offered theoretical and empirical insights to optimize global–local performance.
Federated Network
Food Mass Estimation with LVMs
Built a computer vision pipeline to estimate food mass from RGB images using large vision models (LVMs). Fine-tuned Meta's SAM2 for segmentation and GLPN for monocular depth estimation.
Tools Used:
- Python, PyTorch, Torchvision, Pillow, Transformers, SAM2, GLPN, ViT
Key Achievements:
- MSE: 5.61%, MAE: 1.07% for mass estimation.
- Segmentation MAE: 0.0086, IoU: 95.30%.
- Outperformed prior methods in accuracy.
Research Focus & Innovation:
- Applied alpha compositing to merge SAM2 and GLPN outputs.
- Used a ViT-based model for mass prediction as a foundational step toward accurate calorie estimation.
Food Mass Estimation Pipeline
Remote Maternal Health Monitoring App
Developed a mobile-integrated AI app to monitor and predict maternal health risks in real-time using wearable sensor data, with an MVP prototype for gestational diabetes (GD) prediction.
Tools Used:
- Python, FastAPI, WebSockets, Firebase, PyTorch, Scikit-learn, Transformers, Pandas, Numpy, JavaScript, React Native
Key Achievements:
- Prototype app flagged high-risk GD cases with 95.87% sensitivity.
- Achieved real-time performance with a processing time of 586 ms.
- Achieved an overall prediction accuracy of 95.10%.
Research Focus & Innovation:
- Built a robust AI pipeline with z-score normalization, smart feature selection, ADASYN data balancing, contrastive learning, and FT-Transformer for high-accuracy maternal risk prediction.
System Architecture
Smart Parking Prototype
Built an AI-powered smart parking platform that integrates computer vision, IoT, and multi-platform apps to automate vehicle entry, reservation, and payment processes.
Tools Used:
- Python, OpenCV, YOLO, TensorFlow, Keras, Flask, uWSGI, Nginx, Docker, Docker Compose, React Native, Google APIs, NodeJS, ExpressJS, AWS (EC2, RDS), MySQL, Arduino, Raspberry Pi
Key Achievements:
- Delivered a working end-to-end prototype with > 95% accuracy in license plate and vehicle size recognition, enabling real-time reservations and seamless transactions.
System Architecture:
- Included cloud-hosted AI model, hardware-software communication via microcontrollers, and cross-platform user interfaces (mobile and desktop).
System Architecture
Bird Call Classification
Built a CNN-based model to classify 40 bird species from audio recordings using signal and image processing techniques. Applied Per-Channel Energy Normalization (PCEN) to enhance model performance in noisy outdoor environments.
Tools Used:
- Python, PyTorch, Transformers, Librosa, Scipy, Torchaudio, OpenCV
Key Achievements:
- Achieved 83% multi-class classification accuracy with stable performance across diverse noise conditions.
Key Challenge:
- Extracting reliable audio features from real-world, noisy field recordings while maintaining sensitivity to bird vocalization patterns.
- Accurately distinguishing between similar-sounding species in noisy field recordings while preserving subtle frequency characteristics.
Classification Pipeline
My Research
My Awards
IEEE GLOBECOM 2024 Best Paper Award
I'm proud to share that a paper I co-authored, "Generalization vs Personalization: A Trade-off for Better Data Heterogeneity Impact Mitigation in Federated Learning," received the Best Paper Award at IEEE GLOBECOM 2024. Selected from over 2,200 submissions, it was recognized as the top contribution in the Communications Software and Multimedia Symposium. This recognition highlights our research on improving model performance and ethical AI practices in decentralized learning systems. Grateful to my collaborators and the IEEE community for this honor.
Let's Build Something Amazing
Looking for an AI engineer and full-stack developer who can architect intelligent end-to-end solutions, from model design to deployment? Let's discuss how I can bring your project to life.