Fruit

Designed and specified end-to-end data pipelines, including PostgreSQL schema design, embedding management with pgvector, and data flow across Kafka, Flink, and Airflow.

Developed a full image-processing pipeline: fruit detection with YOLO, crop extraction, classification using MobileNetV3 / ResNet, aggregation, and generation of ripeness metrics and data-driven alerts.

Built a real-time defect-detection pipeline using Flink, including data routing, model processing, and real-time output generation.

Developed API services in FastAPI to serve ML models, pipelines, and data/metrics access.

Implemented an advanced monitoring stack, including Grafana dashboards, custom Prometheus metrics, and system/model performance visualization.

Set up local development and runtime environments using Docker and Docker Compose, including Dockerfile creation and full environment orchestration.

Developed automated tests (Unit and E2E) ensuring data-flow reliability across all system components.

Integrated all core system components (PostgreSQL, MinIO, Kafka, Prometheus, Grafana, MQTT, and GUI) into a coordinated Docker Compose environment, ensuring a fully synchronized and operational system stack.lessly.

YOLO Image Processing: Implemented a pipeline to fetch images from cameras to MinIO, applied the YOLO model to detect and crop images of cows, and stored both cropped images in MinIO and related metadata in PostgreSQL.

Automated Alerts: Processed summary tables of cow maturity data weekly, and automatically sent alerts via MQTT when predefined thresholds were exceeded, notifying relevant systems in real time.

Prometheus Monitoring & Grafana Dashboard: Collected cow maturity and disease metrics using Prometheus, then built a Grafana dashboard to visualize disease and maturity data in real time, including camera-specific statistics.

Server Maintenance and Updates: Maintained the server, ensuring it was regularly updated with Git repositories and new versions of components, keeping the system reliable and up-to-date.

threshold Ripeness Baseline (Python, NumPy) generating ripe/unripe/overripe classifications, weekly rollups, and safety quality-flags for low-confidence predictions. Trained and fine-tuned a MobileNetV3 model (PyTorch) on orchard imagery, including inference pipeline, evaluation metrics (Accuracy, Recall, ROC, Confusion Matrix), and model validation for field conditions.

Developed and deployed a FastAPI inference microservice exposing the ripeness model for periodic batch evaluations and integrated the service into the platform’s Airflow DAGs for scheduled processing.

Implemented a custom Prometheus Exporter in Python for PostgreSQL/pgstatstatements, exposing metrics such as pgvectorquerylatencyms, hnswcachehitratio, and disk usage, and built the “Vector Overview” Grafana dashboard used across teams for monitoring performance and index behavior.

Built a complete Prometheus Docker image including configuration, local runtime fixes, and team-wide metrics onboarding support.

Designed and implemented an interactive Field Visualization GUI enabling map-based device exploration, image browsing per device, and visualization of ripeness predictions for every captured image.