spec · yousentiment-ai · shipped · 2024
YouSentimentAI
End-to-end MLOps pipeline for YouTube comment sentiment
- Role
- Sole engineer
- Domain
- MLOps
- Year
- 2024
- Status
- shipped
stack →LightGBMTF-IDFMLflowDVCAWSFastAPI
A reproducible training-to-serving pipeline: data and models versioned with DVC, experiments tracked in MLflow, and a web app that pulls a video's live comments and charts their sentiment.
01 · Problem
A model that only exists in a notebook cannot be retrained, compared, or trusted. The goal was the pipeline, not the accuracy number.
02 · Approach
- 01DVC for dataset and model versioning; every run is a commit you can check out and reproduce.
- 02MLflow tracking for parameters, metrics, and artifacts across LightGBM + TF-IDF configurations.
- 03Model registry stage transitions gate what reaches the serving app.
- 04Front end fetches comments through the YouTube API and renders sentiment distribution and timeline.
03 · Outcome
- Any past experiment is reproducible from a single commit hash.
- New model versions ship through a staged registry, not a manual file copy.