01The starting point
We built the website for Hundehalterschule Hamburg with our own photos and videos. The next step was the work behind the website, and bookkeeping took the most time. That is where we started.
Case Study · Hundehalterschule Hamburg · AI & Automation
Over 60% less bookkeeping. Live in 2 weeks.
We first built the website for Hundehalterschule Hamburg with our own photos and videos. Then we automated the bookkeeping, with AI on the school's own mini PC.
At a glance
- Client
- Hundehalterschule Hamburg
- Services
- Website with photo & video, then AI & automation
- Hardware
- own mini PC with a GPU (12 GB VRAM), around €2,000
- Result
- over 60% less bookkeeping effort
How the two automations run
Run #0128Done in 5.1 s
Trigger
New invoice
Inbox · IMAPReceived: invoice.pdf
AI agent
Understand receipt
reads, sorts, decidesReading invoice …Sender and amount found✓ Category assigned
- Model
- Inbox
- Bookkeeping
Check
Clear?
Duplicates and amountsChecking …✓ Yes, clearNo, unclear
Automatic
Receipt booked
in the bookkeeping✓ Booked and filed
Human checks
Approve by email
when unclearWaiting for reply …
Receipt bookedSample invoice · dog supplies · €89.90
Question sent by emailReceipt waits for approval by email
Run #0217Done in 3.4 s
Trigger
Payment received
Business accountReceived: course fee
Rule
Calculate shares
VAT and 15% reserveCalculating …✓ Amounts set
Sub-account
Tax
for VAT✓ VAT transferred
Sub-account
Reserves
buffer for later✓ 15% transferred
Reserves set asideVAT and 15% of the amount are in the sub-accounts
02Taking stock
In the first meeting we looked at which programs were in use, how sensitive the data was and which AI model would fit. After that it was clear the school needed its own hardware, enough for up to three people in daily use and with room for more AI agents.
03Its own AI server
We bought a mini PC with 64 GB of RAM and a GPU with 12 GB of VRAM for around €2,000. It runs Ollama for the AI model, n8n for the workflows and Open WebUI as the chat. The model is Qwen3 30B, which this hardware handles well.
04Two workflows for bookkeeping
We automated the receipts first. An AI agent reads every incoming invoice, sorts it and files it in the bookkeeping. If a receipt is unclear, the workflow asks by email and waits for approval. Once that ran reliably, the second workflow followed. On every incoming payment it transfers the VAT and a reserve of 15% of the amount to two sub-accounts, one for tax and one for reserves. We built both in n8n and adjusted them in trial runs until the results were right. A third workflow keeps the server up to date, triggers updates and writes logs.
05Two agents in the chat
Two agents work in Open WebUI and search the web through SearXNG. One follows changes in law and legal topics that affect dog schools. The other is Anna's day-to-day assistant, with access to her inbox, the bookkeeping software and other programs the school uses.
06Operations
We monitor the workflows continuously and check security and pending updates once a month.
07Technology
4 Tools
Automation & AI
2 Tools
Model & hardware
Qwen3 30BLanguage model, runs locally
Mini PC64 GB RAM, GPU with 12 GB VRAM, around €2,000
3 Tools
Connected systems
LexwareBookkeeping
QontoBusiness account
Email inboxIMAP as trigger, SMTP for approvals
Honestly speaking
The hardest part was the hardware. It had to run a large model for three people without the costs getting out of hand.
The agent for legal topics is a first port of call and does not replace legal advice.
Photographed and filmed by us
Testimonial video for the dog school
Let’s connect
Sounds like your topic?
Tell us what you're working on. In 30 minutes, you'll know how to move forward.
Or email ushello@levisteria.com
$ get-in-touch --free
✓ Duration: 30 minutes
✓ Cost: none
✓ Outcome: an honest take on whether it's worth it
# You talk directly to the founders.
→ hello@levisteria.com
