Problem
Move beyond notebook-level ML by deploying a reliable API that can handle real traffic, maintain low latency, and scale correctly in Kubernetes.
Back to projects
MLOps Coursework Project
End-to-End PyTorch Sentiment API
Built for MIDS W255, this project packages a DistilBERT sentiment model behind a production-style FastAPI service with Redis caching, containerized deployment, Kubernetes orchestration, and performance monitoring under sustained load.
Approach
Implemented FastAPI endpoints with typed request/response schemas, baked model artifacts into the container image, added Redis caching, and deployed to AKS with service routing.
Outcome
Delivered a full MLOps lifecycle project with automated tests, containerized inference, and load-test observability using k6 and Grafana.
System Components
Inference Service
- FastAPI app for batch sentiment prediction requests.
- PyTorch + Hugging Face DistilBERT inference pipeline.
- Pydantic models for strict input/output contracts.
- Unit tests with pytest for endpoint and schema behavior.
Platform Layer
- Docker image with model artifact included at build time.
- Redis caching to improve throughput and repeated-request latency.
- Kubernetes deployment on Azure AKS with service routing.
- k6 load tests and Grafana dashboards for performance analysis.
Performance Artifacts