Overview
Fintastic is an AI-driven Financial Planning & Analysis (FP&A) platform that helps businesses automate budgeting, forecasting, and financial reporting. As their customer base grew, the data-intensive AI workloads running on AWS ECS began straining the infrastructure. Resources were statically provisioned for peak loads, the team had no visibility into per-service costs, and deployments were slow and risky. They needed a modern platform that could scale dynamically with AI inference demand while giving them real control over cloud spend.
The Challenge
ECS served Fintastic well in their early stage, but as their AI workloads grew the limitations became clear. Task definitions were manually maintained, scaling was reactive rather than predictive, and cost attribution was impossible at the service level. The team was spending more time managing infrastructure than building their AI models. Peak-hour provisioning meant over 60% of compute sat idle during off-peak periods.
Fintastic was running fixed-size ECS clusters provisioned for peak AI inference loads. During off-peak hours, over 60% of compute capacity sat idle. Deployments required manual task definition updates and took 45+ minutes with no rollback strategy.
Our Solution
Dcode designed and executed a zero-downtime migration from ECS to EKS. We migrated services incrementally, running ECS and EKS side by side behind the same Application Load Balancer, validating each service before cutting traffic over. The entire infrastructure was codified using Terraform with reusable modules.
On the new EKS platform, we implemented ArgoCD for GitOps-based deployments with automatic rollbacks, KEDA for event-driven autoscaling tied to actual queue depth and API request rates, and Karpenter for intelligent node provisioning that matches instance types to real workload requirements. Helm Charts standardized service deployments across all environments.
For observability, we deployed a full stack: Logz.io for centralized logging, Prometheus and Grafana for metrics and dashboards, and OpenTelemetry for distributed tracing across microservices. AWS Secrets Manager secured sensitive configuration. FinOps practices included ECR lifecycle policies to control image storage costs, ALB consolidation via ingress groups to reduce per-service load balancer expenses, VPC endpoints to eliminate NAT Gateway data transfer charges, and AWS Budget Alarms for proactive cost anomaly detection.
This engagement followed the AWS Well-Architected Cost Optimization pillar. Karpenter's consolidation policies, KEDA's scale-to-zero capability, and ALB ingress groups collectively eliminated over-provisioning while maintaining performance SLAs during AI inference peaks.
Results
- 40% reduction in monthly AWS spend through Karpenter right-sizing, KEDA scale-to-zero, and ALB consolidation
- Zero-downtime migration with parallel ECS/EKS execution and gradual traffic shifting
- Deployment time dropped from 45 minutes to under 15 with ArgoCD GitOps pipelines and Helm Charts
- Full observability stack (Prometheus, Grafana, OpenTelemetry, Logz.io) with per-service cost attribution dashboards
- 3x faster autoscaling response with KEDA event-driven scaling matching AI inference demand in real time