Back to work // case study

CareIntake

AI-assisted mobility-equipment intake platform.

status: live in production

My role: I designed and built the AWS infrastructure, CI/CD and deployments. The application code was written by the dev teams.

AWS ECS Fargate ALB RDS PostgreSQL Amazon Rekognition Bedrock (Claude) Bitbucket OIDC ECR

Context

CareIntake is an AI-assisted intake platform for mobility equipment — wheelchairs and related mobility aids. Staff upload patient and equipment photos and complete an intake form; the platform uses AWS AI services to analyze the photos and summarize the submitted form so the record is easier to review.

The problem

CareIntake was a second product that needed the same kind of cloud foundation I had already built for Sentinel: a private, container-based platform on AWS with an automated deploy pipeline. Rather than design the network and delivery from scratch, I treated the Sentinel setup as a reusable ECS blueprint and applied it to CareIntake — then added the piece that makes this product different: an AI layer for the intake flow.

Reusing a proven pattern meant the infrastructure and CI/CD were predictable from day one, so the new work could focus on the Rekognition and Bedrock integration and its permissions rather than on re-deriving a VPC and pipeline.

Architecture

DEPLOY TRAFFIC CONFIG git push Bitbucket CI OIDC → role ECR push image ECS service update Internet ALB public subnet ECS Fargate careintake-app 2 vCPU · 4 GB RDS PostgreSQL private subnet Rekognition photo analysis Bedrock form summary analyze photos summarize intake
deploy / build live traffic config / AI calls
The same ECS spine as Sentinel — ALB to Fargate to RDS, with a Bitbucket OIDC deploy lane — extended with a Rekognition and Bedrock AI layer on the app side.

What I built

Network

  • Same blueprint as Sentinel: a VPC with public and private subnets, the ALB in public subnets and all compute and data in private subnets.
  • The security-group chain (Internet → ALB → app → database) follows the same pattern — see the Sentinel write-up for the full chain rather than repeating it here.

Compute

  • ECS Fargate running the CareIntake app behind the ALB, in private subnets — same blueprint as Sentinel.
  • AI integration (the distinctive part): the app service calls Amazon Rekognition to analyze uploaded photos — wheelchair, mobility-equipment and patient photos — during intake.
  • It also calls an Anthropic Claude model on Amazon Bedrock to read the intake form data and produce a summary, so reviewers get a condensed view of each submission.
  • The ECS Fargate app service runs at 2 vCPU / 4 GB memory per task.

Data

  • RDS PostgreSQL in private subnets, reachable only from the app's security group — same blueprint as Sentinel.

CI/CD

  • Bitbucket pipeline using OIDC to assume a scoped IAM role, build and push to ECR, and update the ECS service — the same deploy lane as Sentinel.

Security

  • Least-privilege task role: the ECS task role is granted only the specific Rekognition and Bedrock actions the app needs to call — nothing broader.
  • Private subnets for the app and database; the ALB is the only public entry point — same blueprint as Sentinel.
  • Bitbucket authenticates to AWS via OIDC, so no long-lived pipeline credentials are stored.

Deploy flow

  1. git push to the CareIntake repository triggers the Bitbucket pipeline.
  2. The pipeline exchanges an OIDC token for a scoped IAM role — no stored AWS keys.
  3. It builds the container image and pushes it to ECR.
  4. It updates the ECS service, which rolls the new task into the Fargate app with minimal-downtime.

This is the same commit-to-live chain as Sentinel, reused unchanged.

Security decisions

Scope the task role to only the AI actions used

The app calls Rekognition and Bedrock, so the ECS task role grants only those specific actions — not blanket AI access. This keeps the AI layer's blast radius tight even though it is the newest and least-proven part of the system.

Keep compute and data in private subnets

The app service and RDS run in private subnets with the ALB as the only public entry point — the same posture as Sentinel, carried over deliberately rather than relaxed for a new product.

OIDC instead of stored pipeline keys

Bitbucket assumes a scoped role via OIDC for every deploy, so there are no long-lived AWS credentials sitting in the CI configuration.

What's next

  • Pin the exact Bedrock (Claude) model version in the task configuration, so model updates are deliberate and reviewable rather than implicit.
  • Add a cost guardrail around the Rekognition and Bedrock calls, as on Sentinel, to keep AI spend bounded.

Stack

AWS VPC ECS Fargate ALB RDS PostgreSQL Amazon Rekognition Bedrock (Claude) ECR Bitbucket Pipelines OIDC IAM

Hiring for DevOps / Cloud roles? Get in touch.

Get in touch
More case studies: Sentinel PetCare MailCheck EC2 vs ECS