We train AI models on your company’s own work. They beat frontier models at your tasks, and you own them.
InvoiceNet 30
Harbor & Pine Supply Co.
Harbor & Pine Supply Co., Portland, Maine
VendorINV-20417
Invoice numberNet 30, due December 4, 2026
Due datePallet racking, installation and freight
$18,240.00 USD
Total dueA fictional invoice. Each field found and tagged, ready for your books.
First Notice of LossHomeowners
Property damage claim
HO-558120
Policy numberOctober 12, 2026
Date of lossWater from a burst pipe under the kitchen sink soaked the floor and lower cabinets overnight.
Cause of loss$6,400 for flooring and cabinet repair
Claim amountA fictional claim form. Each field found and tagged for the adjuster.
Master Supply AgreementArticle XIV
Assignment; Change of Control; Governing Law
Neither party may assign this Agreement without the other party’s written consent.
Anti-assignmentIf Supplier undergoes a Change of Control, Customer may terminate on notice.
Change of controlThis Agreement is governed by the laws of the State of Delaware.
Governing lawA fictional contract. Each clause found, quoted exactly, tagged by type.
Legal is our first case study. The same method works for any repetitive document work: claims, invoices, filings, medical records.
§ I. The problem
They’re trained to do everything, and trained hardest on code. Your work is contracts, claims, accounting, writing: the same task, done thousands of times, by your rules.
§ II. The proof
We trained a model to find specific clauses in contracts, then tested it against Claude on 102 contracts it had never seen.
The test
Each model read commercial contracts and had to find 10 kinds of clause, quoting the exact text. Lawyers had already marked the right answers. None of the 102 test contracts were used in training.
Accuracy here means how often the model quotes exactly the right text. 1.0 is perfect.
The result
It beat Claude Opus by +0.048 and Claude Haiku by +0.102. Both gaps are bigger than the margin of error.
Harder contracts
On a harder set of contracts, Claude Opus fell from 0.659 to 0.620 and Claude Haiku from 0.615 to 0.551. Ours went from 0.681 to 0.694.
A trained model knows how your team marks a clause. A general model has to guess.
Callout
Better prompts didn’t close the gap. Training did.
Also
Every answer comes back as valid, structured data your systems can read.
Harder contracts: they dropped, we didn’t
Accuracy, test set → harder set
| Model | Test set | Harder set |
|---|---|---|
| Our model | 0.681 | 0.694 |
| Claude Opus | 0.659 | 0.620 |
| Claude Haiku | 0.615 | 0.551 |
Tested on 102 lawyer-labeled contracts our model never saw in training. Our model is fine-tuned from an open-weight model.
§ III. How it works
We find where your AI spend and errors actually are.
We build a test from your real data, so we measure what matters to you.
We train an open model on your work, and prove it beats your current model before you commit.
It runs in your cloud or on-prem. You own the weights.
§ IV. Why own it
It learns how your team labels, decides and words things.
Your data stays yours. Nothing goes to someone else’s model.
No surprise updates or shutdowns. What you tested is what runs.
You run your own model, so more volume doesn’t mean a bigger bill per call.
§ V. Where this works
§ VI. A note from the founder
I ran the test on this page myself. Every number here comes from it, and the method is above for anyone who wants to check.
If your team runs the same task thousands of times a day, I’d like to see if we can do the same for your work. The first call is a conversation, not a pitch.
[FOUNDER BIO — I’ll fill this]
— [FOUNDER NAME]