Runtime inventory
Every model-serving process across approved hosts: model, variant, runtime, GPU, precision, KV-cache settings, ports, and container or Kubernetes placement.
// Pilot
AI Runtime Inventory & Drift Pilot
2–4 weeks · read-only · your environment · fixed fee. Know what model is running where, on which GPU, at what precision — and whether it matches what you intended.
// deliverables
What you get
Evidence-grade artifacts you keep, whether or not you continue with the platform.
Drift report
Actual state compared against your intended configuration, machine by machine: what matches, what drifted, and what nobody knew was running.
Fleet dashboard
A single view of the estate for platform and leadership teams.
Benchmark evidence export
Reproducibility evidence tying results to the exact runtime, GPU, precision, and configuration that produced them.
Findings readout
Configuration risks, unknown endpoints, and a governance baseline with concrete next steps.
// safety posture
Designed to pass your security review.
The pilot tooling is read-only by architecture, not by policy.
Read-only, by architecture
The scanner cannot kill processes, stop containers, block ports, or modify hosts. Command execution is restricted to an allowlisted read-only layer.
No sensitive collection
No prompts, responses, secrets, environment variables, model weights, or customer data. Command-line secrets are redacted before anything is persisted.
Runs in your environment
Local scans plus SSH remote scans with key-based authentication only. Nothing leaves your network without your review.
// how it runs
Two to four weeks, one part-time engineer on your side.
Your time commitment: one platform engineer part-time, plus a one-hour readout.
Week 1
Scope hosts, security review, deploy the read-only CLI, first inventory.
Week 2
Intended-state definition with your team; drift analysis; dashboard live.
Weeks 3–4
Evidence exports, findings report, executive readout.