Writing quotes, sending invoices, sorting receipts, answering enquiries: an AI agent can prepare much of this, and you only approve it. I build these systems for freelancers, startups and small businesses. You know the price before we start. The finished system runs in your own tools and belongs to you.
Every project ends with a system running in your tools. It comes with documentation, so you can run it without me.
A client call, an email or a PDF becomes a quote, an invoice or a bookkeeping entry. You approve it before anything goes out.
Agents that work directly with your software through MCP and APIs, and n8n workflows for hand-offs that are still done by hand. When something is unclear or fails, a person gets it.
Do you have an idea, or a prototype that never left the notebook? First we work out what is worth building. Then I turn it into something that runs reliably and is cheap to operate.
After a client meeting, a lot of time still goes into paperwork. meet2invoice reads the call transcript, picks up the agreed scope and price, and creates the client, quote and invoice in Qonto. Language and VAT match the client, including EU reverse charge and small-business schemes.
Each invoice contains a SHA-256 hash of the approved quote, so you can check later that what you billed is what was agreed. Nothing is sent until you confirm.
A small calculator for freelancers who sell packages over several months. It applies VAT to the whole package instead of each month and finds monthly prices where the gross total ends nicely, for example on .99. The app also explains why that makes a one-cent difference.
VATPricingOpen sourceSilverblock brings together DeFi positions, rewards and transactions across more than 20 blockchains. I'm part of the team and built the Supabase database with row-level security, along with the AI agent and the automations.
SupabaseRow-level securityAI agentsAutomationAn open-source hub for MCP servers that can load more than 10,000 tools at runtime. It is published on PyPI and npm and listed in the official MCP Registry.
View on GitHubAn AI assistant in WhatsApp that remembers earlier conversations. It calls tools through a custom MCP server, and its memory is a knowledge graph.
LangChainLightRAGMulti-agent workflows that took over processes which used to be manual: processing documents, scoring leads, enriching data and sending notifications.
n8nRAGWebhooksScope and price are fixed before I start. While I build, you see the progress every week and can try the system yourself.
We talk about your tools and the task that costs you the most time. If AI isn't the right fix for it, I'll tell you.
You get a written offer with the scope, milestones, timeline and price.
There's a working version in a shared repository from the first week. I test it with your real data, including the awkward cases.
You get the code, documentation and a walkthrough. Further support is available if you want it.
Kaiser Data & AI is me. You talk directly to the person who builds your system.
I came to LLM systems from physics research and have spent the last three years building ones that people use every day. I'm stubborn about numbers: if I can't back one up, I don't return it. And I don't trust a green test until I've run the whole thing from scratch myself. A passing test is not evidence. It is evidence that a test passes. In 2025 and 2026 I won nine AI hackathons, including ElevenLabs, NEAR AI and AI Hardcore. Qonto selected my invoicing skill, meet2invoice, for its AI marketplace. I'm also part of the team behind Silverblock, a DeFi portfolio tracker, where I built the database and the AI automation.
Often, yes. If someone on your team spends several hours a week on the same admin tasks, a focused automation usually pays for itself within a few months. We can work out whether that's true for you on the intro call.
A single automation or a proof of concept takes two to four weeks. I split larger projects into milestones, and you can use something after each one.
Each milestone has a fixed price that we agree before work starts. Support and hosting afterwards are optional and billed monthly.
Where possible, your data stays in your own accounts, and only what the task needs is processed. The systems are designed for GDPR, and I'm happy to sign a data processing agreement.
Yes. I build with common tools such as n8n, Python and open APIs. With the code, documentation and walkthrough, you or another developer can carry on without me.
An MCP server connects a program such as Qonto, your accounting tool or your database to an AI assistant, so the assistant can read data there and take actions. A skill is a set of instructions for the assistant: it describes step by step how to do a task, for example turning a call transcript into an invoice. An agent combines both and works through a task largely on its own, with your approval at the important points. An n8n workflow fits when a process is always the same and should run without a chat. It's often a mix. We work out which on the intro call.
Mainly Python, TypeScript, n8n, Supabase and Postgres, plus MCP to connect AI assistants. I choose the model to fit the task and your requirements, for example Claude, GPT, or an open model running on your own hardware if data must not leave your premises. I build inside the tools you already use, not next to them.
Yes, that's how most projects start. After the intro call and the offer, you have a concrete plan with a price.
A few sentences are enough. I usually reply within one working day, and then we set up a free 30-minute call.