Services · Managed Run & Optimisation
Models change, prompts drift, costs creep, integrations move, regulations evolve and operations expose edge cases the test set never covered. Managed Run is the practice of keeping these systems honest, fast, accurate and affordable. Every day, not just on launch day.
The systems we build →What it covers
Health checks, uptime, latency, error rates, queue depths. The boring, essential plumbing: measured, alerted, and on-call when it matters.
The eval harness we built at launch - or that we’ll build if you’ve inherited a system without one - runs on every change and on a schedule in production. We watch for quality regressions in plain language as carefully as we watch for HTTP 500s.
Frontier models change every few months; open-weights models change every few weeks. We track what’s available, test it against your evals, and recommend - and apply - upgrades that move quality or cost in the right direction. Prompts and configurations are versioned, reviewed and rolled out with the same care as code.
Token spend, infrastructure cost, vendor fees: visible in one place, trended, and managed. We cut AI run-costs by routing, caching, batching and right-sizing models. We share the dashboard.
When something breaks, you have one number to call and one team accountable. Post-incident, you get a real review: what happened, why, what we changed to make it less likely next time. We don’t hide behind vendors.
A live backlog of small, high-value improvements, based on the metrics, on user feedback, on the evals, on what we and your operations team are seeing. We run it like an engineering backlog, not like a steady-state support queue.
What we run
If it’s part of how AI does work in your organisation, we can take it on.
How it works
We inventory the systems, the integrations, the data flows, the runbooks (if any), the eval suites (if any), and the people who care about each service. We close obvious gaps before we go live: telemetry that’s not wired up, missing runbooks, evaluations that don’t exist, secrets that are everyone-can-read.
We work to an agreed SLA, with a single accountable owner on our side. Weekly metrics and backlog review with you, monthly business review with sponsors. Quarterly architecture review.
This isn’t a hands-off support contract. We expect to be improving the service every month, and we publish what we did.
Service levels
| Tier | Target uptime | Critical-incident response | Quality eval frequency |
|---|---|---|---|
| Standard | 99.5% | 4 business hours | Weekly |
| Business critical | 99.9% | 1 hour, business hours | Daily |
| Mission critical | 99.95% | 30 minutes, 24/7 | Continuous |
Typical tiers. SLAs are agreed per service during onboarding.
In plain words
We’ll take a look. Tell us what you’ve shipped, where it’s running and what’s keeping you up at night.