Where we know the ground — and what AI does there.

Single agents with a clear role, scoped tools and hard limits — and teams of agents that hand work to each other. We run such a team ourselves.

• Coordinator agent that hands work to specialist agents; independent pieces fan out to parallel worker lanes that return structured results with quoted evidence.
• Per-task routing — who handles it, which model tier, reasoning on or off — decided by two independent judges; any doubt escalates to the stronger model.
• A second, different model checks public output before it goes live; if the check fails, nothing is posted.

Fixed test sets and release gates, so a change to an agent or a model is a release with evidence.

• Regression gate for every model swap: 43 deterministic cases, no LLM as judge. Pass = score of at least baseline minus 2, zero tool-calling regressions, zero request errors.
• Routing benchmark of 60 realistic requests with fixed reference labels, measured before operations cases were added: 95% correct handler, 90% correct model tier.
• Evaluation sets built from your own cases. Our open dataset of 3,413 synthetic agent change requests and a LoRA adapter on Qwen3-4B are public on Hugging Face.

Models on your own hardware when data must not leave your premises: chosen, quantised, served and tested. We run local models ourselves.

• Quantised builds of open models with measured serving profiles — a 125-billion-parameter mixture-of-experts model runs on one Mac with about 40 GB of memory.
• Hybrid by design: parallel worker lanes run on a local model, harder tasks go to a stronger cloud model; a routing decision takes about 1.5 s on local hardware.
• Data classified up front: what may go to cloud models and what must stay local.

Assistants that answer from your own documents, with sources — in the languages your teams and customers use.

• Multilingual assistant on our own website: retrieval across every page in five languages, answers streamed with numbered source links in the visitor’s language.
• Guardrails first: the assistant gets no tools and no customer data, and declines off-topic and prompt-injection requests — 5 of 5 test cases held.
• Q&A on runbooks, policies and project documents with source links — on a private model if the data must stay in-house.

MCP servers and APIs that give an agent exactly the tools it needs — and nothing else.

• Our own website is an MCP server and an A2A agent: other agents can read the studio overview or any page as markdown and send an inquiry.
• Tools scoped per agent, with hard limits and an escalation path to a person.
• We connect agents to ERP, CRM, documents and APIs — or use a plain workflow where no agent is needed.

Coordinating AI projects from detailed concept to pilot: architecture decisions, sizing, data protection, steering.

• Eval baseline and release plan for every agent before the build starts.
• Risk rating and EU AI Act classification per agent (researched knowledge), with human-in-the-loop design and documentation a German works council can review.
• Model-swap gate as part of the project’s change control.

Capabilities we have built: a planning & release cockpit on a private LLM, and the product logic of a multi-tenant AI SaaS.

• Cockpit that checks incoming emails against the live plan and drafts replies — on a private model.
• Multi-tenant AI SaaS logic: case workflows, evidence packages, escalation rules, usage billing.

ChaRM, quality gates, transports, CTS+ and cCTS.

• Agent that pre-checks transports and change requests against quality gates and drafts the CAB summary.
• Transport-sequencing proposals a release manager approves.
• The same gates for the agents themselves: versioned, tested against a baseline, approved and reversible before an agent change goes live.

Solution Manager, Focus Build, Focus RUN; SAP Cloud ALM as researched knowledge.

• Requirements-to-user-story agent for the ALM backlog.
• Health-check summaries for the next steering meeting.
• Inventory of every agent, model, skill and MCP server — owned and versioned in your own Git pipeline or, as researched knowledge, in SAP Cloud ALM.

SLAs, monitoring, disaster recovery, root-cause analysis.

• Incident triage that correlates monitoring alerts and drafts the root-cause analysis.
• Runbook Q&A on a private model.
• Runbook automation in which the agent prepares the steps and a human approves before anything runs.

Non-SAP to ECC, Oracle to HANA, moves to the cloud, landscape consolidation.

• Data-mapping checks before every migration run.
• Cut-over runbook agent with evidence per step.
• Migration test evidence collected by an agent; go-live stays blocked until it is complete.

Security concepts, authorisations, penetration-test coordination.

• Agent that maps findings to action plans and tracks evidence until closure.
• Authorisation-change pre-checks with four-eyes approval.
• Security for the agents themselves: tool scoping and prompt-injection tests before every release.

Commerce and customer-experience systems connected to the SAP core.

• Order-exception triage between commerce and ERP.
• LLM-drafted customer-service replies with human sign-off.

Release planning, coordination across development, test and production.

• Release-note and test-evidence agent.
• Change-calendar guard against freeze windows.
• Model swaps and prompt changes go through the same release gate as code.

Planning, coordination of partners, providers and global teams.

• Agent that checks emails against the live project plan and drafts replies.
• Status reports assembled from the plan, tickets and mail.