AI Agent Development

How Serenica Digital shipped a production agent platform without an infrastructure team

Serenica Digital built, deployed, and shipped Cormac, a contract-first CRM agent platform, entirely inside Orion. Four services, two locked-down agents, one platform.

4 services

Production services on one Orion project: a trust-layer backend, a web surface, and two locked-down Hermes Agents

Under 1

CPU core for the whole platform, about 650 millicores and 2 GiB of RAM

Developer desk at night with a compact server and city lights

The Company

Serenica Digital set out to build a CRM for small businesses that already live in spreadsheets. Instead of building another app, they built an agent. A client brings the spreadsheet their business runs on, a Hermes Agent interviews them inside an Excel task pane, and that spreadsheet becomes a published, versioned semantic contract. Day to day, a second agent takes plain language, like "just closed the deal with Carter," and turns it into proposed changes the user confirms with a tap.

The product is called Cormac, and every change runs through one pipeline: capture, proposal, confirmation, atomic apply, append-only audit. One writer, locked down.

The Challenge

The hard part was not the agent design. It was doing all of it, build through security audit, without spinning up infrastructure somewhere else. A multi-service platform normally means standing up clusters, wiring CI, managing secrets, and building an auth layer before the product work starts. Serenica wanted the whole lifecycle, development, deployment, and production runtime, on one platform they did not have to operate.

The Solution

Cormac runs as four services on a single Orion project: a Node and TypeScript trust-layer backend that is the only thing allowed to write business records, a React web surface, and two Hermes Agent profiles from one runtime. The authoring agent keeps skills on, because the client interview is a skill, with strict write-approval gates. The operations agent runs with skills off, terminal off, and typed tools only, because it handles data access.

Deployment is three vanilla Helm charts, no custom CRDs and no platform-specific operators, packaged as a native Orion plugin with two Secrets as the only external inputs. The same charts run unchanged on a local kind cluster, on Orion, and on EKS. Orion is a swappable substrate, not a lock-in.

Everything lives in one Orion project namespace: pull secrets created before any workload, runtime secrets injected at deploy time and never hardcoded, and workspace-scoped agent tokens, hashed at rest and revocable.

The Results

The complete multi-agent SaaS platform runs in roughly 650 millicores of CPU and about 2Gi of RAM. Serenica's production numbers: an authoring interview costs about fifty cents at a high prompt cache hit rate, and a day-to-day capture costs about a cent and a half at an 82% cache hit rate.

This is not a demo. Cormac ships with settled architecture decisions, typed plugin tools, workspace-scoped tokens, row-level isolation, and an audited security control register. Code to production on one platform, with no infrastructure complexity beyond what is in the repo.

AT A GLANCE

FIELD

AI Agent Development

DEPLOYMENT

Customer-hosted

ABOUT

Serenica Digital is a small product team building Cormac, a contract-first CRM agent platform for small businesses that still run their operations on spreadsheets.

Write your own before and after.

Start a Proof of Value on your own hardware, and we’ll measure it the same way.

Write your own before and after.

Start a Proof of Value on your own hardware, and we’ll measure it the same way.

Write your own before and after.

Start a Proof of Value on your own hardware, and we’ll measure it the same way.