Things I've built.
AI Inspection Tool (SYA)
Built for SYA, a Belgian business that handles trade-ins, purchasing, and reselling refurbished consumer devices. Their workflow required a reliable way to assess incoming devices at scale — without depending on manual inspection for every unit.
Production-ready REST API for automated visual device inspection. Achieved 76% accuracy on inconsistently-labelled real-world data — a non-trivial result given the noise in the training set.
Core design: parallel AI inference pipeline handling side recognition and damage grading simultaneously, with Explainable AI (Grad-CAM) overlays to make model decisions transparent to non-technical stakeholders. Deployed with full CI/CD on Docker and GitHub Actions.
End-to-End MLOps Pipeline
Scalable ML architecture with fully automated retraining triggered on every code push — eliminating all manual intervention in the training cycle.
Cloud-based model training via Azure ML, deployed through Kubernetes microservices (FastAPI behind NGINX reverse proxy), with GitHub Actions orchestrating automatic redeployment whenever a new model passes validation.
TraceWay — Full-Stack AR Navigation
AR navigation platform with crowdsourced risk mapping and hardware sensor integration for real-time mobile safety notifications. Built as a group project.
Work spanned the full stack and full lifecycle: user-centred design (UCD), ERD and data modelling, Vue.js client, Java/Vert.x backend, push notification infrastructure, and quality gates via SonarQube and Vitest.