UptimeTrace

Infrastructure monitoring that keeps predicted risk separate from confirmed incidents, and labels every number LIVE, SYNTHETIC or REPLAY.

Code

UptimeTrace Command Center: host counts by state, a host map, CPU utilisation lines with an experimental forecast, an alerts chart and an evidence-backed insights panel, all marked as preview data. — Command Center running on a generated preview fleet. Every figure is labelled LIVE, SYNTHETIC or REPLAY, and the risk outlook is marked experimental.

[01] problem

On-call engineers need to know what needs attention right now, and a dashboard that blurs real, simulated and predicted data makes that harder to trust.

[02] approach

01Authenticated agents push metrics under a versioned telemetry contract; the backend stores raw envelopes and derives rates itself.
02Each reading is tagged LIVE, SYNTHETIC or REPLAY, and risk predictions are kept structurally separate from confirmed incidents.
03The console is built triage-first: the landing view answers what needs attention now, and absent features show honest empty states instead of fake charts.

[03] what's interesting

The risk model is explicitly experimental: validated only on synthetic simulator data, so the interface never claims accuracy it has not shown.
The language model explains evidence the backend returned; it never sets a risk score or invents evidence.

[04] stack

Python
FastAPI
Go
React
TypeScript
Supabase
Clerk
XGBoost