Projects · 02 · 2021 → now
Projects
Two lines of work: hardware that keeps learning at the edge, and models that learn from unlabeled factory images.
XZ-01 / Edge node · exploded
86 × 22 × 56 mm · blockout
123
Research · 2021 → now01 / 02
Edge device for intelligent IoT monitoring
On-board continual learning
A custom edge-device board that integrates sensors and a microprocessor for real-time IoT monitoring. It supports optimized on-board model training and adaptation, so it keeps learning at the edge without full cloud dependency.
Edge computingContinual learningIoTEmbedded ML
Research · with General Motors02 / 02
Self-supervised learning for welding image analysis
Multi-level contrastive learning
Multi-level contrastive learning methods, MoCo and DINO variants, for GM's large-scale welding image dataset. The multi-scale approach improved representation quality and transfer performance across defect classification tasks.
Self-supervisedComputer visionMoCoDINO
Linked papers2025Unleashing the power of unlabeled plant data: a hierarchical contrastive learning frameworkJ. Manufacturing Systems · 831st2026Harnessing unlabeled plant data and labeled lab data for quality prediction in laser weldingJ. Manufacturing Processes · 157
Case study in progress · waiting on Q2 (GM images)Render · battery module, laser-weld seamsGeneric cell module · no GM data or parts
4