The dozen names of a pork collar
Our back office held every order we ever made and about 1,900 product names. It could not tell us what we sold best, until a small model learned to ask.
JordanOsakaThe line said BONFILE 2-3 kg and nothing else. It sat in one of our order records, typed by someone who knew exactly what they meant. Bonfile is tenderloin; anyone on our desk can tell you that. What no one could tell you, from the line alone, was whether the animal had been a pig or a cow.
At MeatBorsa we run a meat trading desk; at MakeBetter we build tools for the trade. For a long time I believed we knew what we sold. We knew every order. What we did not know was the sum of them.
One product, a dozen names
Our desk runs on BOSS, the back office we use every day. Orders, documents, invoices, trucks, coldstore stock and payments all live there, one record per order. Its margin is not the hopeful margin of the day a deal is agreed. It is worked out after the truck, the insurance and the days in the coldstore, which is to say after the meat has finished costing money.
So the numbers were sound. The words were not. Inside BOSS we had about 1,900 different product names, and every one of them had been typed by a person, from whatever a seller or a buyer had written. Sellers write in their own language. Buyers write in theirs. Neither writes for a report.
Take pork collar. In our records it was Pork neck. It was Schweinekamm, and Karkówka, and karczek, and Karkowka with its accent lost somewhere on the way. It was Collar of pork, and Pork Collar / Neck, and Свински врат. A dozen ways of writing one cut from one part of one animal. A person reads that list and sees a single product. A report reads it and sees a dozen, each of them a thin slice of the truth.
We had a record of every kilo that crossed our desk, and we could not say what we sold best.
The question on Monday
The question we wanted answered is the plainest one a trader can ask: which base products sell best, and at what margin? Not which order. Which product. Every Monday the report was there, and every Monday it answered a different question, namely which spellings sold best.
We did what desks do. Someone kept a spreadsheet, with the names down one side and the real product, as far as they could judge it, down the other. It was careful work, and it was out of date almost as soon as it was saved, because new orders bring new names. Whenever we truly needed the answer, the analysis became a project: a stretch of somebody’s attention, a result, and then the slow drift back to not knowing.
The trouble went deeper than spelling. Often we could not even tell the animal. Bonfile is the tenderloin of something. A tomahawk may be beef, pork or lamb. And “Turkey”, in a trade that crosses borders, is a country as often as it is a bird. A trader squinting at the list could usually settle it, because the trader remembered the deal. The list itself remembered nothing.
The models that had read everything
The obvious thing was to hand the names to the general chatbots, the big general models that will write a contract or a poem on request. We did. They were right less than 55% of the time. They were also slow, and fairly expensive, but those were secondary complaints. A reading that is wrong nearly every second time is not something you can build a report on. You still have to check every line, and now you are checking a machine’s work instead of doing your own.
I do not think the failure was one of size. A model that has read a little about everything has read very little written by people who sell meat by the truck, in a few words, in whichever language was nearest to hand.
A small model that asks
Then we gave the same names to CarcassAI, and the problem was over at once. I would like to describe a struggle here. There was none.
CarcassAI is the product knowledge that sits under every MakeBetter tool. It holds about a thousand meat products, each with one ID. It reads about 7,000 names for them in 18 languages, misspellings included, and it has 26 cutting standards mapped, because the same carcass is divided differently depending on whose standard guides the knife.
It reads a product line twice. First come rules, which are deterministic: the same line gets the same answer on the same release. Then comes Carcass, a small model that MeatBorsa trains itself, on every decision our traders confirm, and it reads the line a second time. Where the two disagree, or where the line is unclear, a person decides, and that decision becomes the next lesson.
Three of its habits matter more to me than its speed. It does not guess: given BONFILE, it asks “which animal?” and waits. It never changes what the seller wrote; the reading sits beside the seller’s words, so the original is always there to check against. And it sends very little away. Only the product phrase travels, with no customer and no price, to a small model on MeatBorsa’s own servers in the EU.
The published figures are these. On sellers it has never seen, it finds the right product 94.6% of the time. One full reading takes a median of 29 milliseconds. The first number is not a hundred, and the design does not pretend that it is. That is what the question, and the waiting, are for.
A report that is never a project
What changed on our desk is simple to describe. Every new order line is read as it arrives, in real time, cheaply and fast. Nobody begins an analysis any more, because the analysis is never finished and never out of date. It is just there. Pork collar is one product again, whatever it was called on the day.
The picture below is a sample, drawn with illustrative data rather than our own figures. Its shape, though, is the thing we could not draw before: sales by animal and then by product, each tile sized by what it sold and marked with its margin.
Illustrative figures. Hover over or tap a tile for its figures.
See the figures as a table
| Animal | Product | Share of sales | Margin | Names in BOSS |
|---|---|---|---|---|
| Pork | Pork belly | 9.8% | 7.9% | 186 |
| Pork | Pork collar | 7.6% | 11.2% | 148 |
| Pork | Pork shoulder | 6.9% | 6.4% | 118 |
| Pork | Pork loin | 5.8% | 8.8% | 97 |
| Pork | Pork leg | 4.6% | 5.1% | 88 |
| Pork | Pork spareribs | 3.5% | 12.6% | 71 |
| Pork | Pork back fat | 2.4% | 4.2% | 52 |
| Pork | Pork tenderloin | 2.1% | 13.9% | 49 |
| Pork | Pork shank | 1.3% | 9.1% | 38 |
| Beef | Beef striploin | 5.9% | 9.4% | 88 |
| Beef | Beef tenderloin | 5.2% | 10.8% | 83 |
| Beef | Beef ribeye roll | 4.8% | 9.9% | 72 |
| Beef | Beef chuck roll | 3.6% | 7.2% | 51 |
| Beef | Beef topside | 3.1% | 6.1% | 46 |
| Beef | Beef brisket | 2.6% | 8.3% | 39 |
| Beef | Beef knuckle | 1.8% | 5.7% | 32 |
| Poultry | Chicken breast fillet | 6.4% | 4.9% | 157 |
| Poultry | Chicken leg quarters | 4.9% | 3.8% | 109 |
| Poultry | Chicken wings | 3.2% | 6.7% | 74 |
| Poultry | Turkey breast | 2.6% | 5.5% | 63 |
| Poultry | Duck breast | 1.9% | 9.6% | 41 |
| Lamb and veal | Lamb leg | 3.1% | 10.4% | 58 |
| Lamb and veal | Lamb rack | 2.4% | 14.1% | 44 |
| Lamb and veal | Lamb shoulder | 2.0% | 8.2% | 39 |
| Lamb and veal | Veal loin | 1.5% | 9.0% | 34 |
| Lamb and veal | Veal shank | 1.0% | 7.4% | 23 |
| All products | 100% | 1,900 |
A chart like this looks ordinary, and it should. That is the strange part of the story. Nothing in it is new information. Every kilo, every invoice, every day in the coldstore was already in BOSS. The only thing missing was agreement about what things were called.
I have come to think that a good deal of what a trading desk knows is hidden in this way: not in data nobody has, but in names nobody has reconciled. A name is not a product. It is what someone called the product on one day, in one language, for one buyer. To the people who typed them, the names were perfectly clear. It was only the report that could not read.
The line still says BONFILE 2-3 kg. We have not corrected it, and we will not; it is what the seller wrote. Beside it there is now a reading, and before the reading there was a question, which one of our traders answered. In all the time we had kept that list, it was the first time the list had asked us anything.