Tapping into a cloud streaming backbone, safety-relevant metadata and events are streamed via Amazon Kinesis Data Streams at thousands of events per second per deployment with sub-200 ms latency, while video processing stays at the edge.

Alerts and response are orchestrated through Lambda and Step Functions, providing real-time dashboards, multiple notification channels (email, SMS, webhooks, mobile), and defined SLAs across Critical, High, Medium, and Low severities.

The system continuously learns from production data, feeding SageMaker training pipelines with a mix of real and synthetic data; active learning and drift monitoring trigger retraining, achieving high-precision and recall on PPE and intrusion models.

The approach shifts from reactive to proactive safety, delivering rapid detection (seconds rather than minutes), 24x7 coverage, and scalable operation across hundreds of concurrent streams while reducing false alerts.

The IRIS platform, built by Tata Elxsi on AWS, enables real-time safety risk detection across many cameras, integrating edge processing, cloud inference, correlation, alerting, and storage.

Security and governance are foundational: a multi-account setup with encryption (KMS), TLS, least-privilege IAM, VPC isolation, and comprehensive logging (CloudTrail/GuardDuty).

Inference runs on SageMaker with dedicated endpoints for PPE, restricted areas, worker safety analytics, and proximity; scales via Application Auto Scaling and micro-batching; each endpoint handles tens of inferences per second with latency under 300 ms.

Key takeaways include edge filtering with metadata streaming, correlation to boost alert trust, ongoing model retraining, and strong privacy governance from day one.

All events and evidence are stored in Amazon S3 with lifecycle management (Standard storage for 30 days, then transition to Glacier/IA), and extracted frames are retained for a year.

Plants face challenges from relying on many cameras, including reactive monitoring gaps, limited human bandwidth, inconsistent compliance, and scalability costs, driving real-time safety gaps.

Event correlation uses Lambda and DynamoDB TTL windows to align detections over 30 seconds to 5 minutes, cutting false positives by roughly 40–50%.

Edge processing trims data by filtering frames at the source (2–5 fps) and sending metadata-only references to the cloud, cutting cloud data by 70–80%, with edge devices powered by AWS IoT Greengrass on NVIDIA Jetson GPUs.