Infrastructure for controlled, auditable AI workflows

Client: Internal product — Dervix. Industry: AI infrastructure.

Running AI-enabled automation for client projects requires more than calling a model API. Every workflow needs consistent handling of usage budgets, human-approval gates for sensitive actions, and a reliable audit log — otherwise each project reinvents the same controls, inconsistently, under time pressure.

The goal was to build this once, as shared infrastructure, rather than as a one-off concern in each project.

The infrastructure needed to support different workflow shapes — scheduled batch jobs, real-time request handling, and long-running agent tasks — without forcing every workflow into the same rigid template. It also needed to make cost and usage visible per workflow, not just in aggregate, so that a single misbehaving process could be identified and stopped quickly.

Budgets, tool-access limits, and approval gates were designed as infrastructure-level primitives that any workflow can declare and rely on, rather than logic each project's engineers have to reimplement. Every action a workflow takes against an external system or a model API is logged with enough context to reconstruct what happened and why, without logging sensitive request content unnecessarily.