Enterprise AI control plane delivered in six months at 75% under budget for a UK software vendor

Client context

swarmd.ai sells a control plane for AI agents to enterprises. When a company wires agents into procurement, recruitment, support or payments, its compliance team wants an audit trail, its security team wants credential isolation, its platform team wants rate limits and kill switches, and its finance team wants a predictable bill. swarmd.ai sits between the agents, the tools and the models and enforces all of that on every hop.

It is a bootstrapped UK software business. The founders funded the build themselves.

The situation before

swarmd.ai had a vision, a deck and a deadline: a real platform in enterprise customers' hands within two quarters, on a fixed budget, in a category whose standards were changing monthly. There was no reference architecture to copy.

The platform had to be a product rather than a prototype: a relay for several agent protocols, a gateway in front of every major model provider, a tool-brokering layer, and a payments rail, all behind one tenant-aware control plane. Policies had to be readable by compliance staff and versioned so an audit could be replayed. Identity had to respect each customer's own single sign-on. And the codebase had to be one the client's own growing team could take over.

What we did

We owned the engineering end to end for six months.

We made the structural decisions in week one and kept them: a modular codebase with typed contracts between every part, an adapter behind every vendor boundary so a new model provider or protocol is an addition rather than a rewrite, and quality gates that block a merge rather than warn about it.

We treated audit and policy as products. Policies are written in a purpose-built language, evaluated by a proven engine, and bound to an immutable version, so the rule that ran on a given day can be shown to a regulator eighteen months later. Every event is written in the same transaction as the business change it records and chained so that tampering is detectable.

We built enterprise identity in from the start: per-tenant single sign-on against customers' own identity providers, central credential rotation enforced everywhere, and encryption of sensitive material before it reaches a database.

We built for handover. Consistent patterns across every service, tests that run against real databases and identity providers, and observability in place before each feature went live. The bar was that a swarmd.ai engineer joining on a Monday would be productive by Friday.

Results

The budget underrun is not only a cost saving. For a self-funded business it is the marketing hire, the months of customer development and the room to say yes to the first enterprise customer who asks for something unexpected.

For the IT reader: technical detail

Structure. A modular monorepo in which every protocol contract, cross-cutting concern and bounded context is its own module; wire formats are defined once and generated everywhere; affected-based CI builds and deploys only what a change touches.

Policy. Customer policies in a purpose-built policy language evaluated by a mature engine, with bindings inherited from tenant scope down to a single agent and each binding stamped with the immutable policy version it was evaluated against.

Audit. Transactional outbox delivery so business writes and audit writes succeed or fail together; hash-chained, tamper-evident events; a hot operational store for live queries and a warehouse for analytics; correlation IDs propagated across every hop.

Identity and isolation. A mature identity broker with custom extensions, per-tenant SSO, central credential rotation, application-layer encryption of secrets, per-service database credentials, and row-level tenant isolation enforced inside the database as well as in code. Every service-to-service call is authenticated with short-lived, audienced tokens.

Pipeline. A typed, ordered middleware pipeline for skill restriction, sensitive-data detection, rate limiting and human-in-the-loop holds, composed by configuration and checked by the compiler.

Stack. Java on the current long-term support release with Spring Boot; Postgres for transactions and ClickHouse for the warehouse; Next.js and React for the management UI; container-native deployment on Kubernetes with rolling, feature-flagged releases; zero-downtime migrations by default.

Practices. Linting, static analysis, security and dependency scanning on every commit; end-to-end tests against real containers on every release; infrastructure as code with reproducible local stacks; structured logging, metrics and dashboards before features go live.