Live demo
Watch an agent investigate an incident
This is the shape of the SRE platform I built at Apple, reduced to something you can run in a browser tab. Pick a scenario and press play. The third one is wrong on purpose — reject it and see what the system does with a bad answer.
p99 latency on checkout-api crossed 2.4s (threshold 800ms) for 5m
service: checkout-api
0/25 steps
The clean path: retrieve, hypothesise, verify against live telemetry, propose, get approval.
Press play to run the investigation.
Human-in-the-loop gate
Every consequential action passes through here. Accept and reject decisions are the evaluation set.
Agent precision
no decisions yet
Runs entirely in your browser — no model call, no backend, no network. The traces are authored to mirror the shape of the real system; service names and numbers are invented. space play/pause · → step · 1–3 scenario.
Grounded, not inferred
Retrieval expands through a service knowledge graph before it touches embeddings. Every claim in the drafted RCA carries a citation ID that resolves to a real artefact — a runbook section, a postmortem, or a telemetry window.
Verification can send it backwards
The verify node is allowed to refute the top hypothesis. When it does, the graph returns to hypothesis with the refutation as new evidence rather than committing to a story that merely sounds right.
The gate produces the eval set
No consequential action executes without an engineer. Their accept or reject is stored as a labelled example, which is where the precision number comes from — it exists from day one instead of being retrofitted when someone asks how well it works.