Error rate across 1.9 million frames processed
Throughput gain over the standard-precision baseline configuration
Modeled annual recurring revenue per school served
DefenX
New York, NY
School Safety Technology
Amazon SageMaker AI, Amazon Kinesis Data Streams, Amazon CloudWatch, Amazon S3, Amazon ECR
DefenX is a school safety technology company that uses computer vision to detect weapons in real time across school camera networks. To scale beyond hardware installed on-site at each school, DefenX needed to know whether its detection workload could run reliably and affordably on the cloud. Avahi validated DefenX’s inference workload on Amazon Web Services, identified the most cost-effective compute configuration, and built a live cloud inference pipeline processing camera feed data end to end. The engagement gave DefenX a proven architecture, a continuously validated pipeline, and the cost and revenue models it needs to price a cloud-hosted version of its service.
DefenX builds computer vision technology that detects weapons in real time from school camera networks, helping school staff respond faster to potential threats. Its detection model runs on the YOLOv11 object detection architecture.
DefenX ran its detection pipeline on local GPU hardware installed at each school. Scaling that approach to more schools meant purchasing and maintaining new hardware at every site, which limited how quickly DefenX could grow and made it difficult to offer a simple, centrally managed service.
To offer a cloud-hosted version of its product, DefenX needed to validate whether its real-time detection workload, roughly 400 to 500 inferences per second per school, could run reliably in the cloud, and at what cost. Without that validation, DefenX could not confidently price a managed, cloud-based offering for its school customers, and risked either overbuilding infrastructure or underestimating the cost of running the service at scale.
As a company planning to offer its detection technology as a managed service across many schools, DefenX needed infrastructure it could scale up or down without owning physical servers, along with access to a range of GPU and AI accelerator options to find the most cost-effective way to run its model. AWS’s breadth of managed machine learning infrastructure, including multiple GPU instance families and a purpose-built AI inference chip, gave DefenX a way to test and validate that infrastructure empirically before committing to a production architecture.
DefenX needed a partner who could move quickly through a structured technical evaluation, benchmark real workloads under realistic conditions, and translate the results into a clear cost and pricing model, not just a working prototype. Avahi’s experience helping AWS customers validate and productionize machine learning workloads made it well suited to lead this evaluation and deliver a foundation DefenX could use to scale its offering with confidence.
Avahi ran a structured evaluation of DefenX’s YOLOv11-based weapon detection model on Amazon SageMaker AI, benchmarking inference throughput and latency across multiple GPU instance types. The team also evaluated AWS Inferentia2, a purpose-built AI inference chip, as a potentially lower-cost alternative, testing it against the same real-time performance targets.
Across these tests, Avahi measured throughput and cost at multiple weight precisions and batch sizes, then stress-tested the leading configurations at up to double DefenX’s target camera load. The results showed that a SageMaker GPU instance running at a lower-precision (FP16) setting with a larger batch size delivered the best combination of throughput and cost, improving throughput by more than 30% over the standard precision baseline. Inferentia2 showed competitive latency but could not match the selected GPU configuration’s throughput without a much larger and more complex instance fleet, so Avahi recommended the GPU-based configuration for production.
With the target configuration selected, Avahi built and deployed a live, end-to-end inference pipeline: Amazon Kinesis ingests camera feed data and streams it to a SageMaker AI inference endpoint running the selected configuration, with Amazon CloudWatch providing real-time monitoring of throughput, latency, and error rates. Avahi then validated the pipeline with a minimum 24-hour continuous validation under sustained load, confirming it could run reliably without manual intervention.
Finally, Avahi built a total cost of ownership model that broke down AWS costs per school and identified the point at which a cloud-hosted deployment becomes less expensive than DefenX’s existing on-premises hardware as more schools are added, along with a monthly and annual recurring revenue model DefenX can use to price a managed, cloud-hosted version of its service.
Benchmarked inference throughput, latency, and cost across multiple Amazon SageMaker GPU instance types and weight precisions
Evaluated AWS Inferentia2 as an alternative AI inference chip
Selected and deployed the production configuration with the best throughput-to-cost ratio
Built a live, end-to-end inference pipeline using Amazon Kinesis for camera feed ingestion and Amazon SageMaker AI for real-time inference
Implemented Amazon CloudWatch dashboards for pipeline monitoring
Completed a continuous, multi-day validation run under sustained load
Delivered a total cost of ownership model with per-school cost projections
Delivered a monthly and annual recurring revenue model to support DefenX’s cloud-hosted pricing
The validated cloud architecture gives DefenX a clear, data-backed path to offer its weapon detection technology as a cloud-hosted service, without owning or maintaining hardware at every school.
Selected production configuration delivered 400-500 inferences/second of sustained throughput, an improvement over the standard-precision baseline
Continuous validation run processed 1,892,656 frames and 247,173 endpoint invocations with a 0.000% error rate
Cost model establishes a pricing floor at a 70% gross margin and models annual recurring revenue of $25,000 per school for DefenX’s cloud-hosted offering
DefenX
New York, NY
School Safety Technology
Amazon SageMaker AI, Amazon Kinesis Data Streams, Amazon CloudWatch, Amazon S3, Amazon ECR
DefenX is a school safety technology company that uses computer vision to detect weapons in real time across school camera networks. To scale beyond hardware installed on-site at each school, DefenX needed to know whether its detection workload could run reliably and affordably on the cloud. Avahi validated DefenX’s inference workload on Amazon Web Services, identified the most cost-effective compute configuration, and built a live cloud inference pipeline processing camera feed data end to end. The engagement gave DefenX a proven architecture, a continuously validated pipeline, and the cost and revenue models it needs to price a cloud-hosted version of its service.
DefenX builds computer vision technology that detects weapons in real time from school camera networks, helping school staff respond faster to potential threats. Its detection model runs on the YOLOv11 object detection architecture.
DefenX ran its detection pipeline on local GPU hardware installed at each school. Scaling that approach to more schools meant purchasing and maintaining new hardware at every site, which limited how quickly DefenX could grow and made it difficult to offer a simple, centrally managed service.
To offer a cloud-hosted version of its product, DefenX needed to validate whether its real-time detection workload, roughly 400 to 500 inferences per second per school, could run reliably in the cloud, and at what cost. Without that validation, DefenX could not confidently price a managed, cloud-based offering for its school customers, and risked either overbuilding infrastructure or underestimating the cost of running the service at scale.
As a company planning to offer its detection technology as a managed service across many schools, DefenX needed infrastructure it could scale up or down without owning physical servers, along with access to a range of GPU and AI accelerator options to find the most cost-effective way to run its model. AWS’s breadth of managed machine learning infrastructure, including multiple GPU instance families and a purpose-built AI inference chip, gave DefenX a way to test and validate that infrastructure empirically before committing to a production architecture.
DefenX needed a partner who could move quickly through a structured technical evaluation, benchmark real workloads under realistic conditions, and translate the results into a clear cost and pricing model, not just a working prototype. Avahi’s experience helping AWS customers validate and productionize machine learning workloads made it well suited to lead this evaluation and deliver a foundation DefenX could use to scale its offering with confidence.
Avahi ran a structured evaluation of DefenX’s YOLOv11-based weapon detection model on Amazon SageMaker AI, benchmarking inference throughput and latency across multiple GPU instance types. The team also evaluated AWS Inferentia2, a purpose-built AI inference chip, as a potentially lower-cost alternative, testing it against the same real-time performance targets.
Across these tests, Avahi measured throughput and cost at multiple weight precisions and batch sizes, then stress-tested the leading configurations at up to double DefenX’s target camera load. The results showed that a SageMaker GPU instance running at a lower-precision (FP16) setting with a larger batch size delivered the best combination of throughput and cost, improving throughput by more than 30% over the standard precision baseline. Inferentia2 showed competitive latency but could not match the selected GPU configuration’s throughput without a much larger and more complex instance fleet, so Avahi recommended the GPU-based configuration for production.
With the target configuration selected, Avahi built and deployed a live, end-to-end inference pipeline: Amazon Kinesis ingests camera feed data and streams it to a SageMaker AI inference endpoint running the selected configuration, with Amazon CloudWatch providing real-time monitoring of throughput, latency, and error rates. Avahi then validated the pipeline with a minimum 24-hour continuous validation under sustained load, confirming it could run reliably without manual intervention.
Finally, Avahi built a total cost of ownership model that broke down AWS costs per school and identified the point at which a cloud-hosted deployment becomes less expensive than DefenX’s existing on-premises hardware as more schools are added, along with a monthly and annual recurring revenue model DefenX can use to price a managed, cloud-hosted version of its service.
Benchmarked inference throughput, latency, and cost across multiple Amazon SageMaker GPU instance types and weight precisions
Evaluated AWS Inferentia2 as an alternative AI inference chip
Selected and deployed the production configuration with the best throughput-to-cost ratio
Built a live, end-to-end inference pipeline using Amazon Kinesis for camera feed ingestion and Amazon SageMaker AI for real-time inference
Implemented Amazon CloudWatch dashboards for pipeline monitoring
Completed a continuous, multi-day validation run under sustained load
Delivered a total cost of ownership model with per-school cost projections
Delivered a monthly and annual recurring revenue model to support DefenX’s cloud-hosted pricing
The validated cloud architecture gives DefenX a clear, data-backed path to offer its weapon detection technology as a cloud-hosted service, without owning or maintaining hardware at every school.
Selected production configuration delivered 400-500 inferences/second of sustained throughput, an improvement over the standard-precision baseline
Continuous validation run processed 1,892,656 frames and 247,173 endpoint invocations with a 0.000% error rate
Cost model establishes a pricing floor at a 70% gross margin and models annual recurring revenue of $25,000 per school for DefenX’s cloud-hosted offering
Let’s explore your high-impact AI opportunities together in a complimentary session
