Policy over prompt
What an agent may do is decided by a versioned policy before execution. Prompts can explain rules, but they cannot grant capabilities. Unknown actions fail closed.
AI & Agents
AI that has to survive production.
Assistant systems, agents and MCP integrations for companies that want more than a demo: traceable decisions, explicit approvals and, where needed, fully local, with no data leaving the building.
No slide-deck consulting. I build and run these systems myself, every day, on my own data.
Think first, prompt second
Not every problem needs AI. And almost none needs ten agents.
The fastest way to burn money with AI: unleash an agent armada on a problem a clean workflow would have solved. The second fastest: do nothing out of scepticism. I work with a simple escalation ladder where every step has to prove itself against the simpler one.
- 01Deterministic when possibleRules, workflows and classic logic are cheaper, faster and testable. A lot of what looks like an AI problem is a data problem.
- 02One model, one clear jobA single, well-guided LLM step with verification solves most real AI cases: classification, extraction, summarisation, drafts.
- 03Tools instead of sprawlTool access, RAG and MCP integration once the model needs to read real data and propose real actions.
- 04Multiple agents, when earnedParallel specialists for review, research or analysis are powerful when scope and requirements justify them. As a starting point they are organisational cosplay.
Complexity is not a feature. It has to be earned. And with the right requirements, it is.
The false analogy
Organisational cosplay: agents are not a workforce.
Human companies are organised hierarchically because humans have human limits: nobody can know everything, attention is scarce, responsibility has to be distributed. An LLM does not have these problems. It has different ones: limited context, errors that compound, no state of its own between steps. Rebuild a company out of agents anyway and you solve problems that do not exist while creating ones that did not exist before.
01
Rebuilding a company
CEO agent, managers, departments: the structure feels familiar, so it feels right.
02
The coordination begins
Every line is a handover in prose. Context gets lost, errors travel along, and every station costs tokens and latency.
03
The rebuild
Away with the role-played hierarchy. An LLM does not need colleagues, it needs clear conditions.
04
Clear assignments, isolated contexts
Deterministic code orchestrates. Specialists work in parallel, each with its own context, without a shared group chat.
05
In the end, the check decides
Tests, schemas and policies verify the result. What passes is adopted, with a receipt.
coordination without information gain
- Every handover squeezes knowledge into prose and loses context on the way.
- Errors do not cancel out, they multiply with every station.
- Coordination costs tokens and latency but produces no information.
- Rules and tasks are explicit instead of hidden in role-play prompts.
- Reviews run in isolated contexts: independent judgements instead of groupthink.
- Verification is deterministic, with tests, schemas and policies instead of a manager agent.
Humans need hierarchy because knowledge and responsibility have to be distributed. An agent system needs the opposite: few, sharply cut tasks, explicit state and checks that are not negotiable. Parallel specialists have their place, but as tools with an assignment, not as colleagues with a calendar.
Control is the feature
Agents that are allowed to act need rules nobody can bypass.
The question is never whether a model can do impressive things. The question is what happens when it gets it wrong. That is why I build agent systems on principles enforced in code, not in prompts:
Receipts and a way back
Every action leaves a receipt. Reversible work stays undoable for days; deletions go through quarantine with a waiting period and a restore point.
Approvals where they matter
Sending mail, moving money, changing infrastructure: an agent can propose such actions but never trigger them itself. Approval stays with a human.
Data access before the model
What a model may see is decided before the prompt exists. Deny by default: unapproved data never reaches the LLM in the first place.
How serious is that? In my own agent system an entire LLM integration is disabled until it can be sandboxed as provably as the others. Switching a feature off because the sandbox is not tight enough yet is inconvenient. It is also exactly why you can trust the system.
What I offer
Three ways to bring AI into your company.
AI features in existing software
Assistance, classification, content generation and vision inside your product or internal system, with cost control, fallbacks and guardrails.
Agents & MCP integrations
Agents that work with your real systems: custom MCP servers for your data and tools, orchestration, approval workflows and an audit trail.
Local & GDPR-compliant
When data must not leave the building: local models on your own hardware, local vector search, local speech processing. Cloud only where you decide.
In production
AI that is already working today.
CRM & Operations
AI rule assistance in Custom Flow
Warning rules, decision logic and AI assistance in the production CRM of an international brand, embedded in existing processes, not next to them.
Event commerce SaaS
The AI suite in TicketBuddy
Seating-plan recognition from photo or PDF, image generation, SEO copy and translations into seven languages, a shop assistant with guardrails. With an included quota instead of bring-your-own-key.
From the lab
J.A.R.V.I.S.: the agent I test all of this on.
An assistant with full access to mail, projects, calendar and knowledge. Possible because everything runs on my own hardware and nothing leaves the machine. Not a product but my lab: this is where the patterns are born that go into client systems after proving themselves in daily use.
Code review · QA · Security · Architecture · Debugging · Requirements · Documentation · Root-cause analysis
Runs daily on real client and project data, precisely because it stays local.
LocalModels, vector search and speech on my own hardware
106 toolsin a custom MCP server for Claude, Codex & co.
8 specialistsparallel review agents with clear roles
Policy engineReceipts, undo, quarantine, explicit approvals
Sharing the knowledge
Talks & training that de-hype instead of upselling.
AI 101
60 minutes of fundamentals for teams without ML background: what AI can do, where it fails, and a checklist for responsible use.
AI Overengineering
Why more agents rarely mean more results: the minimum reliable agent loop and the escalation ladder before adding roles.
Agentic Coding
A live demo of what agentic work actually looks like in day-to-day development: verification instead of circus.
New guide
EU AI Act: when must AI images and AI text be labelled?
An EU-wide practical guide for businesses, with fictional image examples, checklist, AI literacy and official sources.
The entry point
AI assessment: know what is worth it within two weeks.
We walk through your processes and you get a prioritised list of concrete AI use cases: with an honest assessment of where AI is the wrong tool, a cloud-versus-local recommendation including GDPR implications, and an implementation plan with estimates. Fixed price, clear deliverable, no subscription.
Request an assessment
If nothing is worth it, the report will say exactly that.