the deep case study
An agentic MLOps platform, built to be measured.
This one is still being built. One claim runs through the whole build: the classic MLOps loop (register, train, gate, canary, promote, roll back) still works when the thing being served is an LLM agent, provided you add a fourth telemetry axis (tokens and cost) and a fifth artifact type (the reasoning trace). Everything below either supports that claim or is cut.
- Lifecycle stops
- 14
- Loop-backs
- 2
- Units
- 10
- Layers
- 4
- Standing cost
- $0
the story, in three acts
Act 1 · Ship
An agent ships like any model: every version goes into a registry and must pass a gate before it moves.
Act 2 · Drift
Production drifts. A canary slice and monitoring on four axes, tokens and cost included, catch it before users do.
Act 3 · Recover
Rollback is a first-class path back to serving, and every run can be proved after the fact from its traces.
The map below tells that story as fourteen stops and ten units. Click any station and it lights up with who owns it and what that unit proves. Or press Take the tour for a guided walk through seven stops.
Watch the walkthrough (45 s, silent)
The guided tour, seven stops through the lifecycle · recorded from this page, unedited.