I'm an MLOps Engineer focused on designing scalable, cloud-native inference architectures and automated CI/CD pipelines. I specialize in transforming experimental models into production-ready microservices using Kubernetes, Docker, and AWS.

I am currently architecting robust ML systems with a focus on:

• Orchestration: Managing containerized applications with Kubernetes.
• MLOps & GitOps: End-to-end pipeline automation, monitoring, and continuous delivery using ArgoCD.
• Microservices: Decoupling monolithic ML code into scalable FastAPI services.
• Infrastructure: Configuring AWS (EC2, VPC, IAM, ECR) and Linux environments for high availability.

1. End-to-End Fashion AI Recommender (H&M Project)

• Cloud-Native MLOps & IaC: Architected a decoupled deployment strategy by provisioning AWS S3 storage via Terraform, removing heavy model artifacts from Docker images. Optimized model deployment via INT8 Dynamic Quantization (ONNX), achieving a ~75% reduction in footprint (100MB+ to 23MB) with dynamic S3 lazy-loading via Boto3.
• Architecture & GitOps: Orchestrated a highly available Kubernetes environment managed declaratively via Helm and ArgoCD. Automated zero-downtime rollouts and utilized multi-stage builds to reduce image sizes by 54%.
• High-Performance API & Reliability: Architected an asynchronous Redis caching layer, slashing inference latency to <2ms. Validated robustness via Locust, sustaining 805 RPS (equivalent to ~2.9M requests/hour) under 2,000 concurrent users with a 0% error rate.
• Vector Search Engine: Integrated Qdrant as a high-throughput vector database to enable semantic similarity search, optimizing persistent storage structures for millisecond-latency recommendation retrieval.
• DevOps & Observability: Orchestrated a full CI/CD pipeline via GitHub Actions for automated AWS EC2 deployments. Implemented a robust observability stack (Prometheus & Grafana) to track real-time P99 latency and system stability. Enforced rigorous DevSecOps practices (Trivy, Black, Isort) and ensured reliability with Pytest.

2. Cloud-Native Inference Engine (NYC Project)

• Architecture: Deployed a decoupled system on AWS EC2 with zero-trust security via IAM Roles and custom Security Groups.
• High-Performance: Achieved an 85x latency reduction (~281ms to ~3ms) and sustained ~290 RPS by migrating to ONNX Runtime and integrating Redis caching.
• DevSecOps & CI/CD: Established a GitHub Actions pipeline with Trivy for vulnerability scanning. Reduced production Docker image sizes by 68% (2.05GB to 650MB) via multi-stage builds.
• Model Engineering: Engineered a resource-efficient Random Forest model (size reduced by 97%, from 1.2GB to 33MB) and utilized MLflow, resolving critical Kubernetes OOM errors.
• Reliability: Validated system robustness via Locust stress testing (1000 concurrent connections, 0% error rate) and implemented real-time monitoring with Prometheus & Grafana.

3. CRM MLOps & Customer Churn Prediction System

• Cloud Infrastructure & IaC: Architected a highly secure cloud environment using Terraform to provision AWS EC2 instances and configure strict Security Groups. Automated the deployment of a version-controlled AWS S3 bucket to isolate ML model artifacts from the codebase.
• CI/CD & Build-Time Artifacts: Engineered a fully automated GitHub Actions CI/CD pipeline. Implemented an enterprise-grade build-time download strategy, securely fetching ultra-lightweight ONNX models (110KB) and preprocessors directly from S3 during the Docker build phase, ensuring a pristine Git repository.
• Microservices & Containerization: Developed a decoupled architecture featuring a FastAPI inference backend, a Streamlit UI for real-time churn analysis, and Redis for robust state management. Optimized Docker image footprints via custom .dockerignore rules and pushed artifacts to GitHub Container Registry (GHCR).
• Automated Deployment & Reliability: Achieved seamless, zero-touch production deployments. Configured the pipeline to automatically connect to AWS EC2 via SSH, pull updated GHCR images, orchestrated services via Docker Compose, and execute automated environment cleanups (pruning orphaned images) to preserve cloud resources.

LinkedIn