Ask the farm. Check the answer.
Six pens, 28 days of feed, lice, oxygen and mortality data. Ask in English, Norwegian or Dutch. The assistant answers from the data, and code checks every number before you read it.
One question, one answer, every number traced.
A site manager does not want another dashboard on Monday morning. They want to know which pen needs attention and why. Pick a question or type your own.
Fjordvik, as the assistant sees it.
The same data the assistant reads. Pick a pen on the map or in the table to see its last four weeks.
| Pen | Biomass | Avg weight | Lice, last count | Feed vs appetite | Waste | O₂ min | Dead fish | FCR |
|---|---|---|---|---|---|---|---|---|
| t | g | adult female / fish | 7-day avg | 7-day avg | 7 days, mg/L | 7 days | 28 days |
A thin layer on top of systems you already run.
- 1Sources stay where they are
Feed cameras, oxygen sensors, weekly lice counts, mortality registration, the production system. Bought from suppliers, run by them. I would not rebuild any of it.
- 2Facts, not raw dumps
Code turns the data into named facts per pen:
lice.latest,overfeed_pct.avg7,o2_min.min7. The model reads those plus the daily table. - 3The model answers and cites
Claude writes the answer and must list every number it used with its fact key. It suggests actions per role and says when data is missing.
- 4Code checks every number
Each cited value is looked up in the dataset. A wrong or invented number is flagged in red before anyone acts on it. That is the part that makes it usable on a farm.
- 5People decide
Treatment, feeding and harvest decisions stay with the site manager, fish health and the vet. The assistant prepares; it never orders.
What this demo is, and what it is not
It is a working pattern: question in, answer with checked numbers out, in any language your staff use. It is not a computer vision model. Lice counting and pellet detection under water are mature products from specialist suppliers, and that is where they should stay.
- For production read-only access to the data platform, an audit log of every question and answer, and a test set of 50 real questions from site staff that every change must pass.
- Same pattern, other data it works the same on sales and market data: orders, price lists, customer questions. That is where Digital Solutions sits.
- Built in a day with Claude and Cloudflare Workers. The live assistant has a daily limit, so please do not script it.
Peter Heijnen
I run a fish processing company in the Netherlands and I build AI agents and automations every day. I know what a Monday morning looks like when the numbers come from five systems that do not talk to each other.
- Fish, from the inside
- Owner of Heijvis Visverwerking BV, a Dutch fish processor with staff, suppliers and margins to watch.
- AI in daily use
- A team of AI agents that handles mail, bookkeeping prep, websites and reports, built on Claude, n8n, Supabase and Cloudflare.
- Where I fit
- Turning existing data into answers and actions people trust. Not training camera models.