Tata Elxsi swapped a purely recording‑based CCTV setup for a serverless, edge‑augmented pipeline that pushes only filtered frames and structured metadata to AWS, enabling safety‑risk detection in seconds instead of minutes. Engineers care because the pattern demonstrates how to combine Greengrass, Kinesis, SageMaker and Lambda to achieve low‑latency, scalable computer‑vision workloads while limiting bandwidth and operational overhead.
What changed?
Traditional plant monitoring relied on human operators watching recorded video, leading to detection delays of 15–45 minutes. IRIS introduces continuous, automated analysis: edge devices run a lightweight model to discard irrelevant frames, then forward only the frames that contain people, vehicles, or equipment. Those frames are stored in Amazon S3, while a lightweight event containing the S3 key and context travels through a serverless stream for inference and alerting.
Why it matters to AI, cloud, DevOps and security engineers
- Latency reduction – Real‑time inference replaces post‑event review, turning a minutes‑long detection window into seconds.
- Scalable cost model – By filtering at the edge, network and storage usage scale with the number of safety‑relevant events rather than raw camera volume.
- Operational simplicity – Serverless services (Kinesis, SageMaker, Lambda) remove the need for a dedicated video‑processing fleet.
- Compliance support – Structured metadata and S3‑based archival simplify audit trails without retaining full‑resolution video.
Architecture and implementation pattern
The solution consists of two logical layers.
Edge acquisition and pre‑filtering
GPU‑equipped industrial servers (e.g., NVIDIA Jetson AGX Orin) run AWS IoT Greengrass. Greengrass components pull video via RTSP/ONVIF, extract 2‑5 fps, apply motion‑based filtering, and run a first‑pass model to identify frames with people, vehicles, or equipment. Frames that pass are written to an Amazon S3 bucket partitioned by camera, date and hour. Greengrass also handles secure device communication and model updates delivered as S3‑hosted components.
Cloud processing pipeline
A streaming event containing the S3 object key, camera ID, plant, zone and an NTP‑synchronized timestamp is sent to Amazon Kinesis. Downstream, a SageMaker endpoint performs the final computer‑vision inference. The inference result triggers an AWS Lambda function that correlates detections across cameras and time, generates high‑confidence safety alerts, and writes both alerts and the original frame to Amazon S3 for compliance storage. The entire stack runs in the Asia Pacific (Mumbai) region to satisfy data‑residency and latency requirements.
Operational and security implications
Edge devices are managed centrally via Greengrass, which provides mutual TLS authentication and encrypted data transport, reducing the attack surface of the camera network. Decoupling image data from metadata limits bandwidth consumption and isolates raw video from downstream services, simplifying access‑control decisions. Serverless components shift operational responsibility to AWS, but practitioners must monitor Lambda concurrency limits, SageMaker endpoint scaling, and Kinesis shard capacity to avoid throttling under burst loads. Choosing a region close to the plant addresses both latency and regulatory residency concerns.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Adopt a similar edge‑first filtering strategy when dealing with high‑volume video streams to keep network and storage costs predictable. Leverage Greengrass for secure, OTA model distribution and use metadata‑only events to drive downstream serverless processing. Align your region selection with latency and data‑sovereignty requirements. Finally, instrument Lambda, SageMaker and Kinesis metrics to ensure the pipeline scales without manual intervention.