Solution
Caylent collaborated with VMT to design and implement an advanced machine learning solution built using Amazon SageMaker AI, tailored to their unique needs. This initiative combined state-of-the-art AI modeling, MLOps automation, and cloud-native architecture enhancements to maximize efficiency and scalability.
AI Model Training & Optimization
Caylent replaced the existing Rekognition model with an open source model, performing experiments with DETA (DEtection Transformer with Assignment) and DETR (DEtection TRansformer), two object detection models trained on Amazon SageMaker AI. These models were selected due to their ability to handle multiple object classes with improved precision and recall. At the end of experimentation, the chosen model was DETA, for its superior and more reliable detection quality.
To address dataset inconsistencies, Caylent implemented data augmentation techniques, enhancing model generalization and accuracy in real-world moving scenarios.
MLOps, Automation & CI/CD
Caylent built a comprehensive MLOps pipeline to automate model training, evaluation, deployment, and rollback. This included:
- A training pipeline that facilitated automated training, evaluation, and version control using SageMaker AI's capabilities such as Training Jobs, Processing Jobs, Pipelines and Model Registry.
- A deployment pipeline, leveraging SageMaker AI's Real-Time Endpoint, enabled seamless rollouts with A/B testing and rollback mechanisms to ensure stability.
- Endpoint auto-scaling dynamically adjusted resources based on demand, to improve efficiency.
- Manual retraining triggers, which allowed VMT to refine models with new data, ensuring continuous improvement and adaptability.
To ensure smooth and continuous model updates, Caylent integrated Laravel Forge for CI/CD, aligning with VMT’s development environment and tool choice. This enables their teams to iterate faster and reduce downtime.
AI Business Workflow Integration
Beyond model training, Caylent designed a holistic AI solution that integrated directly into VMT’s business operations. This ensured SLA commitments were met while enhancing the overall customer experience. Additionally, Caylent implemented a customized Label Studio platform, enabling VMT to curate high-quality training datasets, further improving model accuracy over time.
Cloud Architecture & Scalability
To ensure cost efficiency and long-term sustainability, Caylent designed a scalable cloud-native architecture, leveraging AWS-managed services for optimal performance. Key components included:
- Amazon SageMaker AI for model training, inference, and MLOps automation.
- EventBridge, Lambda, and S3 for event-driven processing and data storage.
- IAM policies for secure access control and compliance.
- Terraform for Infrastructure-as-Code (IaC), enabling repeatable and efficient deployments.
- ECR-based containerization to support optimized model training and inference workloads.
By implementing these enhancements, VMT’s AI-driven logistics platform is now highly scalable, cost-efficient, and easy to maintain, ensuring the company can seamlessly evolve and adapt to future technological advancements.