Shadow Sentinel

Shipping a new model is risky because you cannot tell how it behaves on real traffic until it is already serving users. Shadow Sentinel removes that risk: a challenger model runs on live requests in parallel with production, affecting no one, while the system measures agreement and decides automatically whether the new model is safe to promote.

• 01On every push, GitHub Actions runs unit tests, code quality checks, and model validation on accuracy and latency, then builds and pushes a Docker image. A failing gate blocks the deploy and notifies the developer.
• 02A FastAPI gateway routes all requests to Model A in production, which serves the user response, while Model B, the challenger, runs silently in parallel on the same traffic.
• 03A comparison layer logs the agreement rate between the two models across requests, plus a confidence gap and high confidence conflicts.
• 04After 200+ requests, the system decides automatically: 80% agreement or higher promotes Model B to production, anything lower triggers an automatic rollback, fires an alert, and keeps Model A. A Streamlit dashboard shows it all live.

This is the closest of my projects to how reliability actually gets handled in production. A model that looks good in a demo isn't the same as one you can trust with live traffic, and the gap between them is exactly what shadow mode, automated promotion, and rollback are for: running a new model alongside the current one on real traffic without it making any live decisions, then promoting or rolling back automatically based on how it actually performs.