01
Describe
Write the use case, the industry, the data, and who actually decides.
Private models for regulated work
Grae SI lets companies in any regulated industry build, validate and run private AI models inside their own infrastructure, even offline, with compliance testing built into every step.
Grae SI helps a twelve-person clinic and a global quality team own their intelligence: a private model, a validation record, and a runtime in their own walls, at a cost that does not require a model-platform department. You keep the adapter. You leave with a model card, an evaluation report, and a runtime that does not need to call home.
Grae SI helps teams prepare evidence for a compliance review. It does not certify compliance with HIPAA, FDA, ISO, SOC 2, ICH, DPDP, ICMR, or any other framework.
Customer-owned weights
The adapter and the GGUF build download to you. There is no Grae-hosted model to rent.
Offline runtime
Ollama, vLLM, or llama.cpp in a bundle you can load on a disconnected host.
Evidence, not a certificate
Every step writes a record a reviewer can read. Nothing here is a certification.
From a paragraph to a package
02
Profile
Compliance agents classify low, medium, or high intensity and name the frameworks.
03
Prepare
See what to collect, what to mask, and a readiness score on a sample file.
04
Train
Pick a small open-weight model and a backend. Grae SI writes one job spec for all of them.
05
Validate
Run a suite matched to the tier. Export a report. It prepares a review. It does not certify.
06
Deploy
Download the package you own and run it offline, on your hardware.
Compliance agents, grounded in the frameworks you already name
Each agent says whether its framework applies, needs a closer look, or does not apply, and why. If a language-model key is configured, Grae SI asks it for a narrative and then keeps the more conservative tier. Without a key, the rule agents run on their own. The demo always finishes.
- AI governanceISO/IEC 42001 and the NIST AI RMF
- SecurityISO/IEC 27001 and SOC 2 evidence, not an attestation
- HIPAAPrivacy, Security, and Breach Notification when PHI is in play
- FDADevice, safety, and Part 11 questions when the record is regulated
- ICHGCP and clinical safety when the work is a trial
- DPDPIndia's personal-data law when India and personal data meet
- ICMRIndian biomedical ethics guidance when that research is in scope
Three ways to train. One package at the end.
Open-weight catalog
Small bases with the license printed on the card: Llama, Mistral, Qwen, Phi, Gemma, plus BioMistral and Meditron where the upstream terms allow a download. Restricted licenses are flagged before you start.
Training backends
Local PEFT, Unsloth, and Axolotl. Managed jobs for AutoTrain, Together, Predibase, SageMaker, Azure ML, and Vertex AI. Or a launcher for your own GPU cluster. Same hyperparameters, data, and eval spec.
Bring your own
Upload a model you already have. Grae SI records the license you declare, fine-tunes through the same adapter interface, and still emits safetensors, GGUF, a model card, and the eval report.
Sandbox profiles
Healthcare providers, pharma, biotech, insurance, diagnostics, CROs, banking, and legal. Each profile is a worked example of the model, the controls, and the effort that stays with the buyer.
Pricing
Compare plansProsumer, self-serve
Studio
$79 per seat / month
Shared registry
Team
$490 per seat / month
Your infrastructure
Enterprise
Custom annual agreement