Nhood · Retail real estate
A platform built around four production workloads
AI engineer, platform build and production onboarding
- Use cases in production
- 4
Nhood had four promising AI use cases and no AI platform to run them. My team built the platform from scratch on EKS. I contributed to the platform itself and brought the four workloads into production under the client's security and FinOps standards.
Production onboarding defined the platform
The platform could not be designed separately from its first workloads. Each use case needed a reproducible deployment path, observable runtime behaviour, and ownership of its cloud cost. Those requirements shaped the shared platform work and the onboarding process together.
Argo CD provided the deployment contract. Application and infrastructure changes were reviewed in Git, reconciled into the cluster, and recoverable without reconstructing manual steps. Helm packaged the workload-specific configuration while preserving one delivery model across the platform.
Cost and operations belonged to each workload
Shared infrastructure obscures which product creates a cost increase. OpenCost attributed Kubernetes spend by workload, while centralised logging and load testing exposed runtime behaviour before each production release. The aim was not a dashboard collection. It was a clear owner for cost, capacity, and failure.
The data layer avoided one database per use case
The four workloads needed shared analytical data without multiplying managed databases. The team used Apache Iceberg on Amazon S3 as the common data layer, with SageMaker Unified Studio for governed access. This kept storage separate from compute and avoided creating a dedicated managed database for each use case.
Outcome and ownership
Four AI use cases reached production on the platform during the engagement. That was a team outcome. My contribution covered both sides of the boundary: platform capabilities and the work required to onboard and operate the production workloads.
Stack
- AWS EKS
- Terraform
- Argo CD
- Helm
- Apache Iceberg
- Amazon S3
- SageMaker Unified Studio
- OpenCost
- SQS
- CloudWatch