Building a Cost-Controlled, AI-Ready Fintech Platform for Periculum

Periculum had a complicated history with AWS. An earlier engagement saw the company burn through approximately $100,000 in AWS credits with little internal visibility into what was driving costs. A third party had set up that environment, and when the credits ran out, Periculum had no clear picture of the cost structure they were inheriting. They moved to GCP and rebuilt discipline around spend, operating at roughly $2,000 per month. Returning to AWS was only on the table if the architecture was intentionally designed for cost control from the start.

Beyond cost, the technical demands of their AI roadmap had grown significantly:

• Cost Governance Failure: The previous AWS environment produced no internal cost transparency. Periculum left AWS because they could not understand or control what they were spending.
• Rapidly Growing AI Workloads: The product roadmap included GPU-intensive model training and fine-tuning, retrieval-augmented generation (RAG) pipelines, agent-based AI systems, and multi-model inference serving for fraud detection and credit risk use cases.
• GCP Migration Complexity: All production workloads ran on GCP, spanning Cloud Run, Cloud Functions, Compute Engine VMs, Cloud SQL, BigQuery, Vertex AI, and GCS. Migrating these to AWS equivalent services required careful mapping and a phased approach.
• No Formal DR Strategy: Despite serving financial institutions, there was no multi-region or multi-cloud disaster recovery plan in place.
• Limited Deployment Automation: Container images were manually deployed, and CI/CD pipelines had accumulated technical debt from a lean team moving fast.
• Always-On GPU Cost Risk: Any architecture that provisioned GPU compute on a persistent basis would recreate the cost overruns Periculum had experienced before.

Vontech Group designed and deployed a production-grade AWS architecture using Terraform for infrastructure-as-code, built around three non-negotiable goals: cost predictability, AI readiness, and zero downtime migration.

• Amazon ECS Fargate: Replaced Cloud Run as the compute backbone for containerized microservices. Each service runs in private subnets, auto-scales based on load, and pulls images from Amazon ECR. GitHub Actions and AWS CodePipeline provide fully automated CI/CD, replacing the manual deployment workflows from GCP.
• Application Load Balancer (ALB) with Multi-AZ VPC: Routes external traffic from banks, dashboards, and partner APIs into private ECS services. The VPC is segmented across two Availability Zones with private subnets isolating application and data tiers from the public internet.
• Amazon Aurora PostgreSQL with pgvector: Replaced Cloud SQL as the primary transactional database, running Multi-AZ for high availability. The native pgvector integration means Aurora stores both structured application data and vector embeddings for RAG pipelines within the same managed cluster.
• Amazon Redshift Serverless: Replaced BigQuery for heavy analytical workloads and BI dashboards. Serverless mode eliminates fixed cluster costs and enforces a pay-per-use model aligned with Periculum's cost discipline.
• Amazon S3 as Production Data Lake: Centralized object storage for raw financial events, historical transaction logs, model training datasets, and document ingestion for RAG pipelines.
• Amazon EventBridge, SQS, and SNS: Replaced GCP Pub/Sub with a fully decoupled, event-driven messaging layer. Fraud detection triggers route through EventBridge to SQS consumers and SNS notification workflows, keeping inference workloads asynchronous and independently scalable.
• Amazon SageMaker: Provides the full AI/ML platform including model training, a Feature Store, a Model Registry, and managed inference endpoints. Supports multi-model serving, container-based custom models, and auto-scaling, replacing Vertex AI.
• On-Demand GPU Instances with Auto-Stop Lambda: EC2 GPU instances (g4dn, p3, p4) are provisioned on-demand for batch training and model fine-tuning. A Lambda function automatically shuts down these instances after use, enforced by CloudWatch monitoring. No always-on GPU spend.
• Amazon ElastiCache (Redis): Low-latency caching for session management and inference optimization, reducing Aurora load during peak API traffic.
• Security and Governance Stack: IAM with least-privilege roles, KMS encryption across databases, S3, and logs, GuardDuty for threat detection, Amazon Macie for PII discovery in S3, CloudTrail for audit logging, and AWS Budgets with real-time spend alerts.
• AWS Database Migration Service (DMS): Handled the migration of structured data from Cloud SQL to Aurora PostgreSQL with minimal downtime.