• AIOps & Observability Prometheus, Grafana, ELK/Splunk, CloudWatch, AI-assisted anomaly detection, SLO/SLI & error-budget engineering
• Cloud & Platform Engineering AWS, Azure, GCP, OpenStack, Kubernetes, multi-cloud architecture
• Production Reliability (SRE) incident management, RCA, blameless postmortems, on-call leadership, resiliency patterns
• AI Engineering LLM/SLM inference monitoring, model drift detection, AI model integration, agentic automation

• Automate the toil first, then automate the response
• Alerts should mean something tune for signal, not noise
• Every incident gets an RCA and a follow-up, not just a fix Infrastructure and policy belong in code and version control Small, safe, reversible changes over bigbang deployments

Production-style multi-cloud observability platform for an agentic AI commerce system.

Monitors 4 Kubernetes clusters across AWS, Azure, and GCP, covering 13 AI agents, 7 adversarial "anti-agents," 6 small language models on a GPU serving fleet, and a PCI-scoped payment path entirely defined as Infrastructure as Code.

• Designed a Prometheus + Thanos + Grafana + Alertmanager stack with a two-tier query model: local Prometheus per cluster for pod-level debugging, global Thanos Query for cross-cluster SLO evaluation
• Built 8 Grafana dashboards (146 panels) and multi-window, multi-burn-rate SLO alerting (fast/slow burn, budget tickets) so paging is proportional to how fast error budget is actually being consumed
• Defined agent-specific SLIs with no classical equivalent grounding score, hallucination rate, safety-review coverage, KV-cache saturation to catch failure modes standard RED-metric monitoring misses entirely
• Enforced cardinality and label governance (forbidden high-cardinality labels dropped at ingest, CI-validated recording-rule naming) and PCI-aware payment SLIs that separate system failures from customer card declines
• Delivered fully as Terraform-provisioned, Git-reviewed infrastructure dashboards and alerts are version-controlled, never edited by hand in the UI

Prometheus Thanos Grafana Alertmanager Kubernetes Terraform AWS/Azure/GCP SLO Engineering

An AIOps control loop that reasons about production incidents and fixes them safely.

Ingests Prometheus/Alertmanager alerts, classifies them, and decides on a remediation using deterministic playbooks combined with a local LLM (Ollama) reasoning layer then executes it against Kubernetes through a pluggable executor, with every action guarded and audited.

• Built the alert → classify → LLM-analyze → decide → execute → record control loop as an async FastAPI service with typed, testable modules
• Added safety guardrails: rate limiting, blast-radius caps, dry-run mode, and human-approval gates, so an autonomous remediation system stays trustworthy rather than a liability
• Implemented an incident memory with similarity search so the system recognizes recurring failure patterns and improves its remediation suggestions over time
• Provisioned the underlying platform with Terraform (AWS EKS) and Helm, with linting, type-checking, tests, and Terraform plan wired into CI/CD (GitHub Actions)

Python FastAPI Ollama / LLM Reasoning Kubernetes Terraform AWS EKS Prometheus AIOps

Serverless order-processing architecture on AWS, built infrastructure-first.

A modular serverless system using Lambda, API Gateway, Step Functions, DynamoDB, and SQS to validate, store, and fulfill customer orders with failure handling as a first-class concern, not an afterthought.

• Designed a Step Functions-orchestrated order workflow with a dead-letter queue and failed-orders table so bad orders are caught, retried, and made visible rather than silently dropped
• Authored the entire platform in Terraform, including remote state locking (S3 + DynamoDB) across environments
• Built a CodePipeline/CodeBuild CI/CD pipeline with a Terraform plan stage, a manual approval gate, and an apply stage plus SNS notifications on pipeline state changes
• Wrote an automated test suite covering both successful order flows and DLQ/failure-handling paths, verified against real CloudWatch alarms and dashboards

AWS Lambda API Gateway Step Functions DynamoDB SQS Terraform CodePipeline CloudWatch

End-to-end CI/CD pipeline: Jenkins, Docker, and AWS EC2.

• Built a Jenkins multibranch pipeline that builds and tags Docker images differently per branch (dev vs master → prod) and pushes them to Docker Hub
• Automated deployment to an AWS EC2 instance, triggered on merge to master, with GitHub webhook integration for fully hands-off delivery
• Configured monitoring on the deployment target as part of the same pipeline

Jenkins Docker AWS EC2 GitHub Webhooks CI/CD

Full-stack AI agent product: voice-based, personalized learning experiences.

• Integrated Vapi.ai for real-time voice interactions with AI tutoring companions, built on Next.js, TypeScript, and Convex for real-time backend/state
• Implemented authentication and subscription billing via Kinde, with progress analytics and session history
• Containerized and deployed the platform to AWS ECS Fargate for production hosting

Next.js TypeScript Convex Vapi.ai (Voice AI) AWS ECS Fargate AI Agents