AI Compliance Evidence, Generated Automatically
CDSCO · EU AI Act · FDA · RBI — one tool, every regulation. What took months and lakhs now takes minutes.
| Group | Accuracy | Sensitivity | Status |
|---|---|---|---|
| Age 18–35 | 0.921 | 0.863 | Pass |
| Age 36–55 | 0.908 | 0.841 | Pass |
| Age 55+ | 0.897 | 0.819 | Pass |
| Female | 0.916 | 0.852 | Pass |
| Male | 0.911 | 0.843 | Pass |
From model to regulator-ready report
Three steps. No manual audits. No compliance consultants.
Connect
Point Valiron at your model and evaluation dataset. Any scikit-learn-compatible model or raw numpy arrays.
Analyze
Automated bias detection, calibration, bootstrap confidence intervals, and performance across every demographic subgroup.
Export
Regulator-formatted PDF. Every section maps to CDSCO MDSW §5–8 or FDA SaMD guidance paragraphs.
This is what the report looks like
Real section numbers. Real CDSCO language. No reformatting needed before submission.
| Subgroup | n | Accuracy | Sensitivity | Specificity | Status |
|---|---|---|---|---|---|
| Age 18–35 | 68 | 0.921 | 0.863 | 0.942 | Pass |
| Age 36–55 | 84 | 0.908 | 0.841 | 0.934 | Pass |
| Age 55+ | 48 | 0.897 | 0.819 | 0.921 | Pass |
| Female | 92 | 0.916 | 0.852 | 0.939 | Pass |
Everything a compliance dossier needs
One Python package. Batteries included. No consultants.
Subgroup Bias Assessment
Per-demographic accuracy, sensitivity, specificity with disparity reports across all sensitive features.
Bootstrap Confidence Intervals
1000-resample percentile CIs on accuracy, F1, sensitivity, specificity, AUC. Regulators require error bounds.
Algorithm Change Protocol
Track model versions and validate no regression after updates. FDA PDCP and CDSCO §7 compliant.
Calibration Analysis
ECE, MCE, and reliability diagram data. Know whether model probability outputs are trustworthy before deployment.
Regulation-Mapped Reports
CDSCO MDSW §5–8, EU AI Act Annex III, FDA SaMD — each section maps to actual guidance document paragraphs.
Zero Data Leaves Your System
All validation runs locally. No model weights or patient data transmitted. PHI never leaves your machine.
Built by someone who felt this pain firsthand
"I spent weeks manually computing subgroup metrics and formatting CDSCO tables in a clinical AI project at SRHU. There was no tool that did this automatically. So I built one."
The old way vs the Valiron way
Manual compliance work takes 3–6 months. Valiron takes one function call.
Common questions
No. Valiron runs entirely on your infrastructure. No model weights or patient data ever leave your system. All validation, bias analysis, and report generation happen locally on your machine.
CDSCO MDSW (India), EU AI Act Annex III, FDA AI/ML SaMD, and RBI ML Model Risk — with more coming. Each regulation maps to the actual section numbers and thresholds from the official guidance documents.
The open source library is MIT licensed and always free. A hosted platform with monitoring and audit trail storage is coming — early access users get priority and discounted pricing.
scikit-learn, PyTorch, ONNX, and any model that outputs predictions as numpy arrays. If your model can return y_pred and y_prob, Valiron works with it.
Get early access + a free CDSCO compliance guide
Sign up and we'll send you a step-by-step CDSCO MDSW §5–8 compliance checklist — the exact document structure regulators expect.
Worth ₹0. Yours free. No credit card.