AI Integration
One workflow where AI saves measurable time, taken to production: a Claude agent with custom skills, an MCP server that gives it named tools on your systems, retrieval (RAG) over your own documents, and browser automation where there is no API. The value is in connecting the model to your knowledge and your tools, safely. I run my own business on this pattern.
Quoted per use case at EUR 90 per hour excl. VAT. Model costs are yours and capped. An evaluation set decides whether it ships.
- One use case first
- Agents with custom skills
- MCP server on your systems
- RAG over your documents
One engagement. Clear deliverables.
Picked for measurable hours saved, not for the demo. Invoice intake, lead follow-up, ticket triage, reporting, a question box over your data.
A Claude agent that works in steps with your tools, following written skills: your procedure, your checks, your tone.
Model Context Protocol: named, permissioned tools in front of your API or database, so every assistant can use them and every call is logged.
Retrieval with sources shown in every answer; embeddings built locally when the data may not leave your servers.
Real examples with known answers, what the agent may never do, what needs a person, a monthly cost cap and a documented switch-off.
Fine-tuned or local models (Ollama, Whisper, a trained wake word) for jobs that must be fast, offline or cheap.
- 01
Intro call: where the hours go today, what a good answer looks like, what data may leave the building.
- 02
Evaluation set and a thin prototype against it, with the model picked per step.
- 03
Production build with the MCP tools, retrieval, guardrails, logging and the cost cap.
- 04
Handover, with the runbook and the evaluation set so you can judge later changes and model swaps.
No guaranteed accuracy, no agent with write access to production without a human step, no vendor lock-in: the model is swappable by design.
Terms in short- Amounts excl. VAT, quote valid 14 days.
- Payment within 14 days, acceptance within 7 days.
- Rights pass to you on full payment.
- General terms and NDA as PDF.
The same API serves the Android app, the admin panel and any AI assistant, through one MCP server with named tools.
The workflow, step by step
Every automation I build has the same shape, whether it runs as a Claude agent, as n8n flows or as a queue job in your application. The value sits in the middle: your knowledge and your tools, connected safely.
- TriggerA mail arrives, a ticket opens, a webhook fires, a clock ticks. n8n, a Laravel job or a cron.
- ClassifyA small, fast model decides what this is and whether it is in scope. Cheap, measured on the evaluation set.
- RetrieveRAG: the relevant pages, tickets or rows from your own data, with sources. Embeddings built locally when the data may not leave.
- Act with toolsThe agent calls named tools through MCP: your API, your CRM, your accounting. Read by default, write only where agreed.
- CheckRules it must pass, and a person where the rules say so. Nothing leaves the building without that step.
- Log and reportEvery call audited, cost counted against the cap, a summary to the people who care.
Picked per step, by job. The cheapest that passes the evaluation set wins; the expensive ones are for judgment, not for sorting mail.
| Job | What I reach for | Why |
|---|---|---|
| Sorting, routing, extraction at volume | Claude Haiku 5.5, or a local small model | Fast and cheap; the evaluation set shows when it is good enough. |
| Drafting, summarizing, answering from your documents | Claude Sonnet 5.5 | The balance of quality and cost for most daily work. |
| Long multi-step agent work, code, hard judgment | Claude Opus 5.5 or Fable 5.1 | When a wrong step costs more than the tokens. |
| Always-on listening, offline, on a device | A small model trained for that one job | Privacy and cost: nothing leaves the device, nothing is billed per hour. |
| Speech to text | Whisper, local or hosted | Good enough for most audio; local when the recordings are sensitive. |
| Data that may not leave your servers | Local models through Ollama, local embeddings | Slower and less capable, and the only honest choice for that data. |
| Glueing services without code | n8n flows | Fine for simple triggers and hand-offs; code when logic, tests and retries matter. |
What the terms mean, and where the value sits
The value is not in the model. It is in connecting the model to your own knowledge and your own tools, safely. These are the parts I build with.
- Agent
- A model that works in steps with tools: reads the inbox, calls your API, writes the result, asks a person when unsure.
- Skill
- A written procedure the agent follows for one job, with the checks it must run. Your way of working, made repeatable.
- MCP
- Model Context Protocol, the open standard that lets any assistant call named tools on your systems. One gateway, every assistant.
- RAG
- Retrieval over your own documents, tickets or database, so answers come with their source instead of from memory.
- Embeddings
- Numbers that make text searchable by meaning. The index behind retrieval; built locally when the data may not leave.
- Evaluation set
- Real examples with known answers. The system ships when it passes them, and every change is measured against them.
- Fine-tuning and small models
- Training a small model for one narrow job that must run fast, offline or cheap. I trained a wake-word model for my own voice assistant this way.
- Guardrails
- What the agent may never do, what needs a human, what is logged, and how much it may spend per month.
Picked per job: the best model for the step, the cheapest that passes the evaluation set, local when the data may not leave.
Claude models since 2024. Each step changed what an agent can be trusted with; the build has to be able to swap the model without a rewrite.
Integrations
Systems that talk to each other reliably: bookings, payments, ERP, mail.
Read more → AutomateFull-stack Build
Design, architecture, backend, front end, API, hosting: a new application built end to end.
Read more → AutomateModernization
PHP, framework, database, web server and hosting brought up to date, while the system keeps running.
Read more → AutomateAndroid App
Kotlin and Jetpack Compose, on your API or a new Laravel backend.
Read more →Which task would you hand to an agent first?
Tell me where the hours go today and what a good answer looks like. I reply within one working day with the use case I would start with, free.