Wassim Sghaier

Built an automated quality-control system for fiber-optic cable installation, replacing manual photo inspection with instance segmentation and custom geometric-analysis algorithms. • Designed two custom algorithms — morphological erosion with centroid analysis, and skeletonization with double-BFS centerline ordering — reaching 95.4% F1 on "sanding verification" and 100.0% F1 on "overbending detection" • Built a self-hosted CVAT workflow accelerated with a serverless Nuclio deployment of Meta AI's Segment

Built an end-to-end ML platform for healthcare data, from raw signal ingestion to calibrated predictions. • Automated data upload, cleaning, and preprocessing in a Streamlit app, cutting setup time 70% • Designed hybrid regression-classification pipelines (PyCaret, Optuna, PyTorch, MAPIE), improving R² 50% • Engineered HRV/PPG-derived signal features, lifting accuracy 20% • Integrated MAPIE so every prediction shipped with a calibrated confidence interval

Prototyped an automated review pipeline for trench-scan photos using Claude via AWS Bedrock — a prompt-engineering approach rather than a trained vision model, used to validate the concept before later building a fully trained segmentation pipeline. • Designed a multi-stage prompt pipeline: one prompt checked whether a valid image was attached, then two ran in parallel checking sanding status and overbending — 85% classification accuracy • Debugged AWS pipelines, databases, and staging environments with t

Delivered automated QC classification at 85% accuracy, by building a multi-step, prompt-engineered reasoning pipeline on Anthropic’s Claude via AWS Bedrock. Raised infrastructure uptime 30%, by debugging backend AWS pipelines and staging environments alongside the DevOps team.

Built a real-time multi-task face-analysis system for face recognition, emotion detection, and gender/age estimation on live video. • Built the detection pipeline in Python with OpenCV and Haar Cascade • Parallelized 3 inference tasks with Python multiprocessing and IPC queues • Designed an SQLite-backed data layer for biometric/health data under a unique identifier

Cut new-project setup time 70%, by building an end-to-end Streamlit ML platform from the ground up – dataset upload, cleaning, preprocessing, visualization, and PyCaret-based model training and ensembling, all in a few clicks. Made imbalanced-regression correction accessible with no custom coding, by implementing Label and Feature Distribution Smoothing (LDS/FDS) from a research paper into a one-click pipeline step. Improved model R2 50%, through iterative experimentation across hybrid regression-classifica

Balanced accuracy against inference latency, by researching and benchmarking models for face detection, recognition, emotion detection, anti-spoofing, and age estimation. Built a real-time, non-blocking face-analysis system, by parallelizing three deep learning models with Python multiprocessing.

Deep learning - Image classification platform