شرح موقعیت
ABOUT XANTORY Xantory runs a vertical-farming control platform that plans, drives and records every crop cycle: climate, light, irrigation and dosing across rooms and racks, from a cloud planning tier down to controllers on the racks. Every reading, every relay switch, every recorded stage boundary and every harvest is stored. What the platform cannot do yet is see the plants. As the first machine-learning engineer on the team, the Machine Learning Engineer – Computer Vision will build that capability: cameras on the racks, models that read the crop, and the pipeline that turns what they see into alerts, records and control decisions. The role will also put the sensor and outcome data already collected to work, and will be measured on what runs in production in front of real crops. Role Overview You will own the work end to end. That starts with specifying the imaging setup and building our first labelled datasets on site, then training models for plant counting, growth stages, stress and disease, and harvest readiness, and deploying them at the edge and on the site server. You will also combine vision results with the sensor and yield data we already collect to compare growing recipes on evidence. This is a hands-on, production-focused role based in Dubai with regular on-site farm work. Success is measured by what runs reliably in front of real crops. Key Responsibilities 1. Computer Vision (Core of the Role) Specify the imaging setup for racks and trays: camera selection and placement, lighting under grow-light spectra, capture schedule, and the image and annotation standards that make a dataset trustworthy. Build the farm's first labelled image datasets from our own crops, starting with data collection on site. Train and deploy models for plant detection and counting, germination and growth-stage recognition, canopy coverage, stress and disease indication, and harvest readiness. Choose the right approach per problem (classification, detection, segmentation, anomaly detection) and prove it with evaluation that holds up on new crops and new racks. Deploy inference at the edge (Raspberry Pi-class devices today, Jetson-class where justified) and on the site server, balancing accuracy, latency and hardware limits. 2. Sensor & Outcome Data Combine what the cameras see with time-series data (temperature, humidity, CO₂, PPFD, pH, EC, flow, actuator history, recorded stage boundaries and yields) to explain outcomes against the recipe. Model cycle length and yield per rack against growing recipes and compare recipes based on evidence. 3. Integration & Operation Ship models as services that fit the platform, including Rust, PostgreSQL and Redpanda for data, Kubernetes for deployment, and results surfaced in Sentinel. Version datasets, experiments and models; monitor model performance and degradation; retrain on production feedback. Define the data contracts your models consume with backend, edge and console engineers. Qualifications & Experience 4 years of hands-on computer vision experience with PyTorch or TensorFlow, including at least one vision system taken from your own data collection and kept running. Experience building image datasets from scratch, including quality control, class balance, and the honest handling of small and shifting datasets. Practical experience with cameras and video pipelines, including exposure, colour under artificial light, and calibration. Exposure to horticulture, controlled-environment agriculture (CEA), plant phenotyping or industrial quality inspection is an advantage. Experience with multispectral or NIR imaging is an advantage. Skills & Competencies Solid Python and ML fundamentals, including leakage-safe validation. Comfortable in a production engineering environment: Git, code review, tests, containers, and working against services others own. Edge inference experience (ONNX, TensorRT, NVIDIA Jetson, Raspberry Pi deployments) is an advantage. Familiarity with MLOps tooling (experiment tracking, model registry and monitoring) is an advantage. Knowledge of Rust or Go, MQTT and Kubernetes is an advantage. Clear written and spoken English; able to explain a model's limits to a grower and its interface to a backend engineer. How Success Will Be Measured Production deployment: Number of vision models running in production on live racks, and time from data collection to first deployment. Model accuracy: Detection, counting and growth-stage accuracy validated on new crops and new racks, not only the training set. Edge performance: Inference latency, uptime and resource use on Raspberry Pi / Jetson-class devices and the site server. Dataset quality: Coverage of labelled datasets across crops and growth stages, annotation consistency and class balance. Model health: Detection of performance degradation and time to retrain on production feedback. Operational impact: Alerts and records adopted by growers, and yield / cycle-length insights used in recipe decisions. Work Location: In person