You are building anomaly detection for an observability product. Customers send you 50 million time series: request rates, error counts, latency percentiles, queue depths, business KPIs: at 10-second resolution. They want to be told when something is wrong without configuring a threshold for every metric.
The current product requires users to set static thresholds. Support tickets say the same two things: "I get paged at 3am every Sunday because traffic drops" and "we missed a real outage because the threshold was set too high a year ago".
You have essentially no labels. Occasionally a user clicks "this alert was not useful", and that is the entirety of your supervision signal.
Design the system. Pay attention to what makes an alert useful rather than merely statistically unusual.
Build the architecture on a canvas: place the components, configure them, connect them into a data flow, and write a short reason for each one. The AI reviewer grades your design against a rubric written specifically for this problem.
Why is the standard deviation a poor dispersion estimate for anomaly detection, and what do you use instead?
A customer deploys at 2pm and 200 series shift simultaneously. What does the user see?
You have no labels. How do you know your new detector is better than the old one before you ship it?
Minimum 6 components · needs a wide desktop screen