Business Challenge

A leading FinTech company, experienced rapid growth over two years, scaling to over 150 microservices deployed on Amazon EKS. While this microservices architecture provided agility, it also introduced significant cost visibility and management challenges:

• Inconsistent and bursty workloads led to significant over-provisioning.
• Idle resources, including EC2, RDS, and EBS volumes, remained active for extended periods.
• No real-time insight into service-specific cloud usage or cost contribution.
• Manual optimization cycles took 6–8 weeks and required deep DevOps involvement.
• Existing cloud cost management tools were reactive and lacked AI/automation capabilities.

Objectives

Suvan Infitech was brought in to:

• Reduce Company's monthly cloud spend by at least 25%
• Implement AI-based automation to eliminate manual cost review cycles
• Improve visibility and accountability across business units for cloud usage
• Introduce real-time governance workflows via modern team communication tools

Suvan Infitech’s AI-Driven Solution

Suvan Infitech designed and deployed a robust AI-Powered Cloud Optimization Platform, combining ML forecasting, LLM agents, automation pipelines, and Slack-native governance. The solution had four core modules:

1. Predictive Workload Modeling

   To understand usage patterns and reduce over-provisioning:

   • Implemented time series forecasting models (Prophet, LSTM) to analyze hourly/daily usage trends across EKS pods and EC2 nodes.
   • Automatically predicted compute requirements and matched them to optimal instance families and purchase types (On-Demand, Reserved, Spot).
   • Used real-time metrics ingestion via CloudWatch and custom exporters to continually retrain models.

   Result: Services dynamically right-sized every 6 hours based on real-time usage and forecasted needs.

2. AI-Based Rightsizing & Anomaly Detection

   To detect waste and optimize resource allocation:

   • Built an ML model to analyze historical CPU, memory, disk, and network utilization across EC2, EKS, and RDS.
   • Identified under-utilized and over-provisioned instances; provided actionable resizing and scheduling suggestions.
   • Deployed Isolation Forest models to detect sudden or abnormal cost spikes due to misconfigurations or inefficient queries.

   Result: Reduced idle resources from 15% to less than 3%.

3. Intelligent Autoscaling Policies (RL-Based)

   To replace static thresholds with adaptive scaling:

   • Deployed a Reinforcement Learning (RL) agent trained on workload behavior over 3 months.
   • Automatically adjusted Kubernetes HPA and EC2 Auto Scaling Group policies based on demand curves.
   • Policies learned and adapted to seasonal events (e.g., quarterly transaction load spikes, Black Friday traffic surges).

   Result: Improved scaling responsiveness by 60%, while lowering cost through proactive capacity planning.

4. Cost-Aware Scheduling with GPT-4 Agent

   To optimize non-production workloads and improve decision velocity:

   • Integrated OpenAI GPT-4 via LangChain to build an intelligent agent capable of reading cloud logs, usage data, and cost analytics.
   • Suggested off-hours shutdown schedules for staging/dev environments and batch jobs.
   • Built a Slack-native chatbot interface for real-time recommendations and one-click approvals.

   Result: DevOps teams could approve actions like "shut down staging EKS cluster after 7PM" via Slack/Teams in seconds.

Quantifiable Results

Technology Stack & Tools