Owl Vision

Rebuilt Owl Vision's pest detection system and deployed to 50+ devices across 18 apple farms in New England. The system detects codling moth, obliquebanded leafroller, and oriental fruit moth infestations before visible damage appears—enabling farmers to intervene during the critical first wave, when treatment is most effective and before pest populations multiply exponentially.

Achieved 2x improvement in detection accuracy and 10x faster inference compared to the previous CNN-based system, enabling near real-time alerts to farmers in the field. This resulted in up to 50% less pesticide application and 10% increase in production.

Researched and migrated from image classification to YOLO-based object detection with bounding box localization of pest targets. Designed preprocessing pipeline (intelligent cropping, tiling) to maximize detection of small-scale pest features at high resolution. Architected modular Docker deployment on AWS, replacing legacy EC2 setup with containerized system that decouples model retraining from infrastructure—enabling rapid iteration without deployment rewrites or dependency conflicts.

Coordinated annotation of thousands of training images across multiple contributors, building self-service documentation to scale data collection independently. Collaborated with mobile app developer to align model API outputs (bounding box coordinates, confidence scores) with frontend requirements. Navigated stakeholder management when proposing infrastructure changes—when questioned on migrating to Docker ("if it works, why change it?"), explained production reliability risks of manual deployments: dependency conflicts, version mismatches, and cascading impact of server downtime on farmer alerts. Advocated for containerization as both a stability measure (reproducible deployments) and enabler of faster iteration (safe local testing before production releases). Translated technical concepts for farmers, explaining model behavior (detection confidence, environmental factors) in practical agricultural terms.

Self-directed project timeline and deliverables as the sole ML engineer, balancing model research, data pipeline development, production deployment, and stakeholder coordination. Identified and advocated for infrastructure improvements despite initial resistance, building business case around deployment reliability and development velocity.