Services · Guardrail models
Guardrail models built on your data, small enough for every call.
We build small, dense guardrail models for the risks you care about, from your own data, with the same recipe as our own family. Each is fast enough for the request path and portable enough to run anywhere, and it's yours.
Our guardrail model family
- Personal data, v0.1Finds and masks 77 kinds of personal data in English.Open weights
- Personal data, v0.2Adds XML, SQL, log and key-value traffic to prose.Open weights
- Personal data, multilingualThe same job across languages.Built in-house
- Named entitiesPeople, organisations and places in free text.Built in-house
- ToxicityAbuse and harm in what people and agents send.Built in-house
- Prompt injectionAttempts to override instructions, including disguised ones.Built in-house
Six models, two with open weights on Hugging Face.
The recipe
Small, dense, fast and portable.
Small
Built from scratch for one job, with its own tokenizer and architecture, not a general model cut down. Our open personal-data model is 43 MB.
Dense
One pass over the text gives several decisions at once, each from its own small head on a shared encoder, so a check costs one read, not several.
Fast
Made for the request path, where every call waits for the check: 1.4 ms median per check on an ordinary 8-core CPU.
Portable
Each model is one ONNX file. It runs on a CPU, at the edge or fully air-gapped, and in VidaiGuard beside the Vidai gateway.
The evidence
Measured in public.
*PIIMB English leaderboard, 17 October 2026: Stirling-Minimo PII v0.1 ranked 3 of 36 at 0.904 average masking F2 across four English sets, 0.014 behind the leader, a 1.4-billion-parameter model. Size, speed and coverage from the Stirling-Minimo PII model card; on a GPU it serves 7,600 checks a second on one previous-generation GPU, using about 3 GB of its 24 GB.
Build yours
A model for your risk, trained on your data.
- 1
Define the risk
What must be caught, in which languages and formats, and what a miss costs you. Your policy, not a generic safety list.
- 2
Gather the data
Your own examples, labelled with you, plus generated cases for what's rare. We can work inside your environment, so the data never has to leave it.
- 3
Train from the recipe
A small model built for that one job, from the recipe behind our own family.
- 4
Prove it
Measured on your held-out data and on public benchmarks where one exists, with a regression suite that runs on every retrain.
- 5
Ship it
One ONNX file, running in VidaiGuard beside the gateway or anywhere ONNX runs.
What you own
Everything, with no dependency on us.
The model
The weights and the ONNX file, under your control. No call back to us to run it.
The pipeline
The training and data preparation code, in your repository, so you can retrain as your data changes.
The evidence
The evaluation suite and its results, to show an auditor or a risk committee what the model catches and what it misses.
Where it helps most: regulated data in healthcare, finance and legal work; languages and formats a general model handles badly; and anywhere a check must run without a GPU or a network.
Production systems shipped at
Amazon · GE · Tata
and other operators, not slide decks.
Also
The other ways we help.
Enterprise architecture
Vidai deployed inside your boundary, as code: high availability across regions, and Python middleware moved onto the Rust gateway.
How it works →The sovereignty and cost audit
Where your AI data goes, whose law can reach it, and what ungoverned spend costs, from your real deployment, with a ranked plan to fix it.
Shadow AI audit checklist →How we start together