AI & MLOps
An end-to-end MLOps production infrastructure engineered for automated training, versioning, containerization, and deployment of RNA structure prediction models.
FastAPI
Docker
Kubernetes (k3d)
NGINX Ingress
PostgreSQL
Redis
Azure ML
GitHub Actions
1. Architecture & System Design
The core objective of this project was automating the complete path from model development to a resilient, scalable production environment. The microservices architecture decouples model inference, data management, and caching into independently scalable services.
2. System Implementation
- Model Pipeline in Azure ML: Configured reproducible training pipelines in Azure Machine Learning, tracking experiment metrics and automatically registering artifacts in the Azure Model Registry.
- Inference API (FastAPI): Developed a high-performance backend API that validates input RNA sequence payloads via Pydantic schemas and delivers predicted structure coordinates.
- Caching & Data Persistence: Leveraged Redis as an in-memory cache to serve repetitive prediction queries instantaneously. PostgreSQL stores prediction metadata, query audits, and inference logs.
- Kubernetes Orchestration (k3d & NGINX): Containerized all services with Docker and orchestrated them inside a lightweight local k3d Kubernetes cluster, utilizing NGINX Ingress for traffic routing and load balancing.
- CI/CD via GitHub Actions: Built an automated pipeline triggered on pull requests to run code linting, execute test suites, build fresh Docker images, and handle rolling cluster deployments.
Focus: Reliability and reproducibility — ensuring a zero-downtime, fully automated path from code commit to containerized production rollout.