Your own AI model, in a few clicks.
Classify tickets, orders and documents with a model trained on your data. More accurate than , the market's reference model, it runs on your servers and never bills per use.
- $25 per model
- Beats or your credit back
“Does my PIN come with my card?”
- typestrictGetting the physical card
- Card arrival
Off-the-shelf AI charges a lot for a one-word answer.
It does not know your business
It gets your categories, your terms and your way of deciding wrong.
It bills every question
Millions of decisions a month become a bill that only grows.
It takes your data away
Every question sends your customers' data to a third party.
From spreadsheet to model in four steps.
No code, no GPU and no machine learning team.
Send your spreadsheets and files, or start from a public dataset. Messy is fine.
Pick a base model, prepared by our lab to serve all kinds of businesses.
Our agent organizes your data into standardized datasets, enriches it with synthetic data and public datasets, and anonymizes sensitive data. You review it and, with one click, training starts.
Get your production-ready AI model, trained on your data and tailored to your tasks, beating market models or your credit back.
points above in banking support, over 1,001 tests.
Measured on data, not on slides.
Public datasets run through the product the way a customer would: 20% of each held out for testing and measured on the same examples.
| Benchmark | typestrict | Jev | Gain over Jev | 95% CI |
|---|---|---|---|---|
| Banking supportBANKING77 · 1,001 tests | 0.915 | 0.798 | +11.8 | |
| InsuranceBitext Insurance · 802 tests | 0.991 | 0.896 | +9.5 | |
| Financial newsFinancial news · 804 tests | 0.824 | 0.743 | +8.1 | |
| TelecomBitext Telco · 806 tests | 0.991 | 0.911 | +8.0 | |
| MortgagesBitext Mortgage · 814 tests | 0.987 | 0.920 | +6.7 | |
| TravelBitext Travel · 792 tests | 0.984 | 0.952 | +3.2 | |
| Retail bankingBitext Retail Banking · 806 tests | 0.990 | 0.963 | +2.7 |
Across the 15 tasks we ran: 7 wins, 5 ties and 3 losses to . See them all on the blog
Your model, your rules.
You pay once, run it wherever you want and keep owning everything.
No per-use bill
Pay once. A million decisions cost the same as one.
Runs without a GPU
About 165 MB, answers in milliseconds on a regular CPU.
Wherever you want
Container, Cloud Run, AWS Lambda, Firebase, browser or phone.
No strings to us
The file is yours and keeps running even without typestrict.
Anonymized by default
Sensitive data is masked before training, and the original is deleted.
Commercial licenses only
Every base model and public dataset we use allows commercial use.
One model, one price.
No subscription and no GPU bill. Credits never expire.
- The agent organizes your data
- Model performance report
- Model to download and run anywhere
Training more than one?
In a pack, each model costs less.
Beats on your data or your credit back. Measured on 20% of your data held out for testing.
Questions
What is a base model?
The starting point of your model: a small AI our lab prepared to decide (classify, prioritize, route). In training, it learns your company's way. The technical name is encoder.
What does our agent do?
It reads your files as they come, builds a standardized dataset, fills in with public and synthetic data and masks personal data. You check it before training.
What do I get at the end?
A file with your model in the open ONNX format, which runs from Python, JavaScript, Java, C# and on phones, plus the performance report.
How do I connect it to my company?
Like any service: your system sends a text and gets the decision back as JSON, with the probability of each option.
Where do I run it?
Wherever you already run your systems: a regular container, Cloud Run, AWS Lambda or Firebase Functions, your own server, the browser or a phone. No GPU.
Your first model costs $25.
Start with a spreadsheet. The agent handles the rest.