Deep Light is an offline AI assistive wearable I developed as a capstone project, now with InkyLab AI. It runs YOLO11 object detection, MobileFaceNet face recognition, and ML Kit labeling on-device, with spoken results, fall detection, and an SOS alert.
I trained YOLO to detect object types beyond the COCO dataset and benchmarked models before deployment. During full-system testing, I fixed 13 faults, including a face-distance error traced to an incorrect reference width.
I designed Deep Light around privacy: user pictures aren't saved to external clouds, and face data is stored as embeddings rather than images. After diagnosing overload in an always-on pipeline, I redesigned it for on-demand inference and coordinated background and UI threads so live processing wouldn't freeze the app.
I also built LLM agent workflows for classification, enrichment, routing, confidence scoring, and rule-based escalation, and developed event-driven microservices deployed on AWS with Kubernetes. I'm pursuing a master's degree and hold the AWS Certified Developer certification.
