Novaeon Studio · The AI Agent Factory

Agents,engineered forproduction.

We design, build, release and run AI agents on the hardest stacks — SAP included. With 22+ years of enterprise SAP delivery behind every release.

AI-generated photorealistic image: sunrise over the horizon of the Earth seen from orbit, with the Novaeon station and agent vessels moving along three orbits through four gates

  • Open modelsPublished on Hugging Face
  • Open sourceActive in the OpenClaw community
  • 22+ yearsEnterprise SAP delivery behind Novaeon
  • Release disciplineSAP release & change management (ChaRM)

The problem we solve

A demo agent takes an afternoon. An agent your auditors, your works council and your operations team will accept is architecture: defined role, tested skills, clear limits, a human who signs off — and a release process that lets you change it without fear.

How an agent ships

Idea → Agent → Release → Run.

The same path every time. Nothing reaches production on enthusiasm alone.

  1. 01

    Idea

    Pick the process where an agent pays off. Map inputs, decisions, risks and owners — and decide where a plain workflow is the better answer.

  2. 02

    Agent

    Build the agent and its skills: tool contracts, guardrails, prompt-injection tests, eval scenarios and the human approval step.

  3. 03

    Release

    Promote it like software: versioned, tested against a fixed baseline, risk-rated, approved, traceable — and reversible.

    • Eval gate
    • Risk rating
    • Approval
    • Rollback ready
  4. 04

    Run

    Watch quality, cost per action and drift in production. Improve on a release cadence, not by hot-fixing prompts.

Enter the Factory

What we build

Four things. Built to run, not to impress in a demo.

AI agents

Multi-step agents with one clear role, scoped tools, hard limits and an escalation path to people. Measured by eval scenarios, not by a good first impression.

  • Role & limits
  • Tool scoping
  • Human-in-the-loop

Skills

The deterministic building blocks agents call: look up, validate, create, post. Small, tested, versioned — the part that makes an agent reliable.

  • Deterministic
  • Tested
  • Versioned

Integrations & automation

ERP, CRM, APIs, documents. We connect the systems an agent needs — and automate the work that needs no agent at all.

  • MCP servers
  • APIs
  • Workflows

Private & sovereign AI

Local and self-hosted models when data must not leave your infrastructure. Agents on your hardware, under your control.

  • Local LLMs
  • On-premise
  • EU data

Free download

Free: the release checklist for taking AI agents to production

40 checkpoints in 9 blocks show whether your agent is inventoried, tested, approved, monitored and reversible. A free PDF, no email required.

In-house · Agent Ops

We run the factory ourselves.

Before we build agents for clients, we run them in our own company, with the same quality and release controls we implement for our clients.

  • 43deterministic gate cases per model swap
  • 2 / 8candidate models passed the gate
  • 95%correct handler, 60-task routing benchmark
  • 90%correct model tier, same benchmark

The Novaeon Agent Factory

  • Model swaps became releases: of 8 candidate builds, 2 passed the gate and 6 were stopped before production.
  • One model under test tried to announce a server reboot it had no tool for — it surfaced in the gate, not in production.
  • Routing on our 60-task benchmark: 95% correct handler, 90% correct model tier; a routing decision takes about 1.5 s on local hardware.
Approach
  • A coordinator agent delegates to specialist agents (studio, operations); five or more independent pieces of work fan out to parallel worker lanes that return structured results with quoted evidence.
  • A router makes three decisions per task — who handles it, which model tier, reasoning on or off — with two independent judges; any doubt escalates to the stronger model.
  • A regression gate for every model swap: 43 deterministic test cases, no LLM as judge. Pass = score at least baseline minus 2, zero tool-calling regressions, zero request errors. Each swap now also produces a release record, entered in a hash-chained audit log.
  • Everything public is fact-checked against sources and checked by a second, different model before it goes live.

Our own benchmarks, October 2026. The routing benchmark uses 60 realistic requests with fixed reference labels (measured before operations cases were added).

Hard stacks

Where IT mistakes are expensive.

The enterprise stack we have delivered on — and the agent stack we run ourselves.

Track record — delivered

  • SAP S/4HANA
  • SAP ERP 6.0 / ECC
  • Solution Manager 7.2
  • ChaRM
  • CTS+
  • cCTS
  • SAP HANA
  • SAP BW/BI
  • SAP CRM
  • SAP APO
  • SAP Fiori
  • SAP Hybris
  • Oracle
  • AIX

Our agent stack — in daily use

  • Claude
  • MCP
  • A2A
  • Python
  • TypeScript
  • Node.js
  • Local LLMs
  • Eval harnesses
  • Docker
  • Linux
  • Cloudflare
  • Astro
  • Hugging Face
  • Git-based release

Where we are

Based near Bucharest. Working across Europe.

Novaeon Studio, or simply Novaeon, is a Romanian AI agent studio. Our registered office is in Sat Dudu, Chiajna, Ilfov, directly at the western edge of Bucharest, Romania. We work with companies in the Bucharest area, across Romania and across the EU.

Questions

Before the first call.

Do you only work with SAP customers?

No. SAP is where our release discipline comes from and where agents are hardest to get right, but we build agents, skills and integrations on any stack — including fully self-hosted models.

Our SAP system is on-premise. Can we use Joule?

There is no standard Joule path for on-premise S/4HANA or ECC today; SAP ties it to cloud editions and specific contracts. The readiness check tells you what is possible in your landscape — and which agents you can build without waiting for Joule.

What does “AI release management” mean in practice?

Every change to an agent — prompt, model, tool, data source — goes through versioning, a fixed eval baseline, a risk rating and an approval, with a known-good release to roll back to. Like ChaRM, for agents.

Can our data stay in-house?

Yes. We build on local and self-hosted models when data must not leave your infrastructure, and we scope every tool an agent can reach.

How does an engagement start?

With a 30-minute Agent Strategy Call, then an audit. You decide after the audit whether and what we build.

Do you train your own models?

We don’t train foundation models. On Hugging Face we have published an open dataset of synthetic agent change requests and a LoRA adapter for Qwen3-4B trained on it. We offer dataset creation, evaluation sets and LoRA fine-tuning of open models as a service. Our other public Hugging Face builds are quantised versions of open models, and every model is tested against a fixed gate before it goes live.