Impactful Work/Cloud Cost Optimization & Infra/DevOps Automation (₹222K → ₹185K/month)
Cloud Cost Optimization & Infra/DevOps Automation (₹222K → ₹185K/month)
13 technologies · 9 highlights
What I Built
- Led a platform-wide FinOps and infrastructure optimization initiative across the full Azure + AWS footprint - auditing every VM, disk, managed service, and AI workload against real utilization data - cutting monthly cloud spend from ₹222K to ₹185K (17% reduction, ₹37K/month saved) with an active optimization roadmap targeting ₹150K/month (~32% total reduction) - all with zero performance regression and zero downtime for production traffic.
- Executed a utilization-driven VM right-sizing program across 10+ virtual machines: analyzed CPU, memory, and IOPS metrics per workload over sustained windows, downgraded over-provisioned instances that were burning budget at single-digit utilization, and upgraded genuinely constrained VMs where resource pressure was hurting latency - sizing every machine to its measured requirement instead of its original guess.
- Hunted down and decommissioned every source of orphaned spend: unwanted mounted VMs, unused servers, and deactivated services were fully deallocated and removed from resources completely - including their attached disks, static IPs, and network interfaces - so nothing invisible kept billing after the workload it served was gone.
- Implemented workload-aware storage tiering across managed disks: Premium SSDs retained only where IOPS/latency actually demand them (database and hot-path workloads), while logs, backups, and low-IO volumes were downgraded from Premium to Standard SSD - and under-provisioned hot disks upgraded - aligning storage cost precisely with each workload's real I/O profile.
- Optimized managed service footprints: right-sized and consolidated email services and Redis instances - tier corrections, connection pooling, and TTL/eviction policy tuning - so supporting services stopped running at capacity tiers the actual traffic never justified.
- Enforced AI cost governance at the platform level: introduced usage limits on AI infrastructure and AI services (per-user rate limits capping worst-case inference spend) and optimized context-window and token usage - trimming prompt payloads, capping response windows, and reusing cached responses - so AI cost per request dropped without degrading answer quality, and total AI spend became a bounded, predictable line item instead of an open-ended one.
- Eliminated structurally expensive footprint items: removed the Azure Bastion dependency and high-cost D8 VM instances, and fixed bandwidth cost spikes at the routing/caching layer - attacking the largest single line items on the bill first for maximum savings per change.
- Drove every decision from measured data, not assumptions: utilization monitoring and cost-analysis dashboards per resource group, budget alerts on anomalous spend, and a monthly cost-review loop that treats the cloud bill as an engineering metric - which is how the ₹185K result keeps compounding toward the ₹150K target instead of creeping back up.
- Executed all resizes, tier migrations, and decommissions with production-safe change discipline: changes scheduled in low-traffic windows, health checks validating each workload after every resize, and rollback paths held ready - the entire optimization program shipped without a single user-facing incident.
Tech Stack
Azure Virtual Machines (right-sizing, deallocation)Azure Managed Disks (Premium SSD ↔ Standard SSD tiering)Azure Monitor (CPU / memory / IOPS utilization metrics)Azure Cost Management (cost analysis + budget alerts)AWS EC2 / S3 / CloudFront (footprint audit)Redis (tier right-sizing, TTL/eviction tuning, connection pooling)Email Services (Azure Communication Services / AWS SES optimization)AI Services Governance (usage limits, rate caps)Token & Context-Window Optimization (prompt trimming, response caching)Bandwidth / CDN Optimization (cache-first routing)Bastion Removal + D8 VM EliminationPM2 / Node.js Workload ConsolidationZero-Downtime Change Management (low-traffic windows, health checks, rollback plans)
Key Concepts
- FinOps: Cloud Bill as an Engineering Metric
- Utilization-Driven VM Right-Sizing (10+ VMs, up AND down)
- Orphaned Resource Elimination (VMs, disks, IPs, NICs)
- Workload-Aware Storage Tiering (Premium ↔ Standard SSD)
- Managed Service Right-Sizing (Redis, Email)
- AI Cost Governance - Usage Limits + Token/Context Optimization
- Attack the Largest Line Items First (Bastion, D8 VMs, bandwidth)
- Budget Alerts + Monthly Cost-Review Loop
- Zero-Downtime, Zero-Regression Change Discipline
- ₹222K → ₹185K Achieved, ₹150K Target in Progress
Low-Level Architecture (LLD)
High-Level Architecture (HLD)
Impact
- Cut monthly cloud spend from ₹222K to ₹185K - a 17% reduction saving ₹37K every month (₹4.4L+ annualized) - with the optimization loop actively driving toward ₹150K/month (~32% total, ₹8.6L+ annualized run-rate), achieved through measured right-sizing rather than blanket cuts.
- Right-sized 10+ VMs in both directions - over-provisioned machines downgraded, constrained machines upgraded - so the savings came with zero performance regression: latency-sensitive workloads actually got faster while idle capacity stopped billing.
- Eliminated 100% of identified orphaned spend: unwanted mounted VMs, unused servers, and deactivated services removed completely along with their attached disks, IPs, and NICs - closing the silent-billing leaks that accumulate in every long-lived cloud account.
- Aligned storage cost with real I/O profiles via Premium ↔ Standard SSD tiering - premium performance retained exactly where databases and hot paths need it, and pure cost paid nowhere else.
- Made AI spend bounded and predictable: per-user usage limits cap worst-case inference cost, and context-window/token optimization plus response caching cut the cost per AI request without degrading answer quality - turning the fastest-growing cost category into a governed one.
- Removed the largest structural line items - Azure Bastion dependency, high-cost D8 VM footprints, and bandwidth cost spikes - capturing the highest savings-per-change first before fine-tuning the long tail.
- Institutionalized cost discipline with budget alerts and a monthly cost-review loop - the cloud bill is now tracked like an engineering metric, which is why savings are compounding month over month instead of creeping back.
- Shipped the entire program with zero user-facing incidents: every resize, tier migration, and decommission executed in low-traffic windows with health checks and rollback plans - proving aggressive cost optimization and production stability are not a trade-off.