Skip to content
Pehra

AWS

On AWS, Pehra runs in Mumbai on one Graviton instance, from one CloudFormation template. The stack is built and tested, not deployed yet.

Pehra architecture: the committee in a browser, a phone or laptop, the Pehra bridge and CCTV recorders over RTSP reach one EC2 Graviton4 instance with COOL in AWS Mumbai. On it the web app, the camera processes running OpenCV 5, the AI check with its OpenCV tools and the job worker; flagged moments go to the AI check, incidents to the job worker, and alerts to the society's Telegram group. S3, Bedrock, Parameter Store, CloudWatch, SNS and CloudFormation complete the stack.
The parts and where they run. Open the picture for full size.

What each service does

EC2 on Graviton4
One c8g instance from the COOL image runs everything, with OpenCV 5 built for Graviton and KleidiCV.
Amazon S3
Pictures, clips and a Litestream copy of the database.
Amazon Bedrock
Claude Haiku 5.5 checks flagged moments, compiles watches and answers questions.
SSM Parameter Store
Settings and secrets, loaded when each service starts.
CloudWatch
Logs, metrics, a dashboard and alarms.
SNS
Alarm emails to the operator.
CloudFormation and Budgets
The whole stack from one template, with a monthly budget alarm.

Running the instance is estimated at about $105.00 a month, an estimate from list prices and not a bill, before Bedrock, S3 and data transfer (the estimate).

Deploy it

  1. An AWS account on the Paid plan, and the AWS command line with a profile named pehra.
  2. In AWS Marketplace, subscribe to COOL for Graviton4 and copy its image id for Mumbai.
  3. In the Bedrock console, turn on access to Claude Haiku 5.5.
  4. Point Google sign in, the Polar webhook and your Telegram bot at your domain.
  5. Copy the parameters file, keep it private, and fill in the COOL image id, an alert email and your keys.
  6. Run the deploy as a dry run, read it, then run it.
  7. Point the domain's A record at the address it prints, and open /healthz.
cp deploy/params.env.example deploy/params.env && chmod 600 deploy/params.env
deploy/deploy.sh --dry-run
deploy/deploy.sh

The deploy refuses to finish unless OpenCV from COOL is the one running and the app fits its memory budget. Details are in the README and the template.