Keep the model close to the work.
Train, test and run private AI models on hardware you control.
Get startedsh4pe
Keep the model close to the work.
Your knowledge, your machine, your record.
sh4pe trains an open-weight model on your own documents, tests it on questions it has never seen, and serves it to the tools you already run. It all happens on hardware you control.
Your data stays in the building.
Training, testing and serving all run on your own machine. Your documents, data and models never go to a cloud service.
The model changes only when you change it.
No vendor update moves your results. You pin a version and can reproduce any answer.
Every run leaves a record.
Each training run and each exam writes its files to a folder you own: the recipe, the checkpoints and the score.
sh4pe serving a trained specialist to other apps on the same machine. Every call is logged. Its text never is.
Four steps, one machine.
Bounded dataset
Question and answer pairs come from your own documents. A seal check confirms that no exam question sits in the training set.
The seal check, computed on the machine. Clause text never leaves it.
Local training
A small adapter teaches an open-weight model your material, on your own machine. The recipe, checkpoints and loss curve stay with the weights.
One training run, read from its own folder.
Held-out evaluation
The model sits an exam on questions it never saw. Every answer is graded against an exact key and written to a file.
The score is a file on disk, not a number in a vendor's portal.
Qualified use
Only a model that passes, with a named person's sign-off, goes to work. Other apps then call it like any local model.
The same exam clause, asked twice. The trained specialist returns the exact words.
Measured on a sealed exam.
- 34%
- 76%
- 5 min
Strict match on 203 exam questions the model never saw. Every answer is graded against an exact key.
How it was measured
Measured 25 September 2026 on six public Canadian pension policy documents. Qwen2.5 3B, exam sipp-holdout200-k.
The training recipe and the full exam record are in the technical sheet. Download the technical sheet
Your hardware, set up by DAP.
- Any setup. We fit sh4pe to the hardware you already have.
- Installed and configured by DAP.
- Open-weight models, kept on your own disk.
- Other apps call it like any local model, including Overture.
The models live on your disk. The network is needed only to download one.
Models propose. Humans ratify.
A model takes on a task only after it passes a test on data it has never seen, and a named person signs off. Running locally does not, on its own, give a model that authority.
Chief investment officers, chief compliance officers and general counsel at pensions, asset managers, OCIOs and insurers.
Who uses it, and how
Teams working with sensitive institutional knowledge and explicit data boundaries.