Machine Learning Engineer - Video Analytics

About Ezlo Innovation

Ezlo Innovation is a leading IoT platform developer powering smart home and property management solutions across 60+ countries. Our family of brands—including Vera, MiOS, Fortrezz, and Centralite—brings nearly 50 years of combined experience in home automation and IoT markets. We deliver cloud-to-ground, white-label IoT solutions to security dealers, property management companies, builders, utilities, and retail partners worldwide.

About the Role

We're looking for a Machine Learning Engineer to own and advance our CloudML video analytics platform. This system processes video from smart home cameras to detect objects including people, vehicles, animals, packages, faces, and license plates. You'll be the primary developer responsible for this project while collaborating with our broader cloud infrastructure team.

ML model performance

The Platform

CloudML v2 is a production video object detection service built on:

• ML/Computer Vision

  : YOLO v8/v11, YOLOWorld (zero-shot detection), PyTorch, OpenCV, face-recognition, OpenALPR

• Backend

  : Python 3.11, FastAPI, Redis Queue (RQ)

• Infrastructure

  : Kubernetes, Helm, Docker, NVIDIA GPUs (CUDA 12.6)

• Observability

  : OpenTelemetry, structured logging

The system processes MP4 video files through a multi-stage pipeline: download, validation, frame extraction, multi-model inference (standard objects, packages, faces, license plates, barcodes), and webhook delivery of results.

Primary Responsibilities

Model Performance & Accuracy

• Improve detection accuracy across all object classes (person, vehicle, animal, package, face, license plate)
• Reduce false positive and false negative rates
• Fine-tune and retrain YOLO models on domain-specific datasets
• Evaluate and integrate newer model architectures as they become available
• Develop robust evaluation metrics and benchmarking pipelines

Model Efficiency & Optimization

• Optimize inference speed and resource utilization
• Implement model quantization, pruning, or distillation techniques
• Balance accuracy vs. latency tradeoffs for real-time processing
• Optimize frame skip strategies and batch processing
• Profile and eliminate performance bottlenecks in the inference pipeline

Ongoing Development

• Maintain and extend the existing detection pipeline
• Add support for new object classes as business needs evolve
• Improve hotzone (region-of-interest) detection logic
• Enhance face recognition matching accuracy
• Collaborate with the cloud team on infrastructure and scaling

Required Qualifications

• Strong experience with object detection models

  , particularly YOLO family (v5/v8/v11), and understanding of model training, fine-tuning, and evaluation

• Hands-on experience with PyTorch

   for model development and optimization

• Computer vision fundamentals

  : image/video processing, frame extraction, bounding box handling, NMS, confidence thresholds

• Python proficiency

   with production-quality code practices

• Experience deploying ML models

   in production environments (containerization, GPU inference, batching strategies)
• Understanding of 

  model optimization techniques

  : quantization, TensorRT, ONNX conversion, etc.
• Familiarity with 

  Linux/Docker/Kubernetes

   environments

Preferred Qualifications

• Experience with video analytics or surveillance systems
• Background in face recognition systems
• Experience with license plate recognition (ALPR/OCR)
• Familiarity with zero-shot detection models (YOLOWorld, CLIP-based approaches)
• Experience with dataset curation, annotation, and augmentation
• Knowledge of edge deployment considerations (model size, latency constraints)
• Experience with distributed job queues (Redis Queue, Celery)

Tech Stack You'll Work With

Category

Technologies

ML Frameworks

PyTorch, Ultralytics YOLO, YOLOWorld, OpenCV, Supervision

Specialized ML

face-recognition (dlib), OpenALPR, Tesseract OCR, pyzbar

Backend

Python 3.11, FastAPI, Redis, RQ (Redis Queue)

Infrastructure

Kubernetes, Helm, Docker, GitLab CI/CD

Hardware

NVIDIA GPUs, CUDA 12.6

Observability

OpenTelemetry, structured logging

What Success Looks Like

• Measurable improvements in detection accuracy (precision/recall) across object classes
• Reduced inference time per video while maintaining or improving accuracy
• Well-documented model evaluation pipelines and benchmarks
• Proactive identification and resolution of model performance issues
• Clear communication with the broader team on ML capabilities and limitations

Employment Details

• Type

  : Full-time

• Team

  : Cloud Infrastructure Team

• Scope

  : Primary owner of CloudML video analytics platform
•