MLOps
Scene Mood Classifier: one container, three clouds
Multimodal classifier shipped as an operated MLOps platform
- Timeline
- Aug 2025 to Dec 2025
- Affiliation
- WPI DS-553 Case Studies in Data Science
- Clouds from one image
- 3
- Mood scores
- 5
- Live Spaces
- 2
What it is
A CLIP plus Wav2Vec2 fusion model that scores the mood of a video or image-and-audio pair, shipped as one container image deployed to GCP, AWS and Azure with monitoring, alerting, CI/CD and tested recovery paths.
Problem
Most course projects stop at a model; the goal here was to prove the model can be operated reliably across clouds with measurable recovery behaviour.
Approach
Gradio UI plus an HTTP inference API with per-class probability logging and latency tracking; Dockerized app and API; Prometheus, node exporter and Grafana dashboards; GitHub Actions workflow chaining with Discord deployment status; per-provider port injection with runtime-aware config.
Highlights
VM bootstrap and recovery paths with auto-restart, health checks and Discord alerting, tested against forced crashes, full wipes and network drops with recovery time treated as a performance metric.
Standardized port assignment eliminated an entire class of downtime across providers.
Published as two Hugging Face Spaces (local mode and inference-API mode with Hugging Face OAuth).