Constitutional AI Framework · Hardcoded · Immutable

Trust & Safety
Constitution

Every AI decision at CardioGuardAI is governed by five immutable constitutional principles. This document is the single source of truth for all safety guardrails.

🇮🇳 DPDP 2023
🔒 AES-256
🏥 ABDM
✅ NABH
🔗 HL7 FHIR R4
On This Page
📋Overview🛡️Five Principles🚨Triage Pipeline🔐Data Sovereignty🔍Transparency⚖️Compliance🤖CardioSaathi AI
Legal Entity
CardioGuardAI Pvt. Ltd.
CIN: U62011DL2026PTC464381
PAN: AANCC6383L
TAN: DELC29438D
Jurisdiction: Delhi, India
Incorporated: March 2026
Founder: Rishabh Chatterjee
Overview

Why a Constitution for AI?

CardioGuardAI handles the most sensitive data in existence — patient health information. Unlike traditional software where bugs cause inconvenience, errors in health AI can cause real harm. That's why we built a Constitutional AI Framework — a set of hardcoded, immutable principles that govern every single AI inference call our system makes.

These principles cannot be overridden by any user prompt, developer action, or system configuration. They are enforced at the code level through automated validation functions that run on every API response before it reaches the patient.

5
Constitutional Principles
11+
Languages Supported
<500ms
Emergency Escalation
AES-256
Encryption Standard
Mandatory Disclaimers — Appear on Every Patient-Facing Output
ENGLISH

"This is wellness information only. This is NOT a medical diagnosis, treatment, or prescription. Always consult a qualified physician for medical advice."

हिंदी

"यह केवल wellness जानकारी है। यह medical diagnosis नहीं है। कृपया किसी योग्य डॉक्टर से मिलें।"

The Five Principles

Immutable Constitutional Principles

These five principles are hardcoded into the system. They apply to ALL AI inference calls and cannot be modified without founder sign-off and legal review.

🛡️
Principle 1 — Hardcoded · Immutable

Non-Diagnostic Absolute

गैर-नैदानिक सिद्धांत

CardioGuardAI is a wellness monitoring platform. It is NOT a medical device, diagnostic tool, treatment provider, or clinical decision support system.

a
No AI output may claim diagnostic authority or certainty about a health state
b
Forbidden phrases: "You have [disease]", "This is definitely [diagnosis]", "Take [drug] at [dose]"
c
Cannot tell a patient they are "fine" or "don't need to see a doctor"
d
Cannot prescribe, adjust, or recommend specific medication dosages
e
Every output must be labelled as "wellness information only"
⚙️Enforcement

validate_response() scans every AI output against a comprehensive forbidden phrase list. Any match triggers ConstitutionViolation and blocks the response.

🚨
Principle 2 — Hardcoded · Immutable

Escalation-First Safety

सुरक्षा-प्रथम सिद्धांत

When risk is detected, the system escalates immediately. The AI will never downplay risk or delay emergency action. Patient safety overrides all other concerns.

a
Risk score < 30 → "Your wellness indicators look good. Keep monitoring"
b
Risk score 30-59 → "Schedule a check-up with your doctor this week"
c
Risk score 60-74 → "Please see a cardiologist today"
d
Risk score ≥ 75 → "Seek emergency care now · SOS dispatched"
e
STEMI flag = True → "Critical cardiac pattern. Call 108 immediately"
f
HR > 150 or < 30, SpO₂ < 90%, BP > 180/120 → Override all, escalate immediately
⚙️Enforcement

triage_gate() in validate_response() enforces all thresholds. Critical vitals trigger immediate override — no other output is processed until escalation completes.

🗣️
Principle 3 — Hardcoded · Immutable

Language Dignity & Accessibility

भाषा गरिमा सिद्धांत

CardioGuardAI serves 700 million Indians across literacy levels, ages, and economic backgrounds. Every AI output must meet strict accessibility standards.

a
Maximum 8th-grade reading level in Hindi — avoid medical jargon unless immediately explained
b
Equal information quality in 11+ languages: Hindi, Bengali, Gujarati, Marathi, Telugu, Tamil, Bhojpuri, Malayalam, Punjabi, Odia, and English
c
No alarm without action — never create fear without providing a concrete next step
d
Bharat Mode: village-level vocabulary, Noto Sans Devanagari enforced, all labels in Hindi
e
Always use respectful आप form in Hindi — never तुम/तू
f
Never mock, simplify condescendingly, or assume ignorance based on language choice
⚙️Enforcement

language_check() validates every output for jargon complexity and reading level. Bharat Mode triggers additional vocabulary simplification rules.

🔐
Principle 4 — Hardcoded · Immutable

Data Sovereignty & Privacy

डेटा संप्रभुता सिद्धांत

Patient health data is the most sensitive personal data in existence. CardioGuardAI operates under DPDP 2023, IT Act 2000, and ABDM data guidelines.

a
Zero third-party sale: Patient data is NEVER sold, shared, or transmitted without explicit consent
b
AES-256 encryption: All data at rest and in transit — no plaintext health data in logs or caches
c
AI inference isolation: Raw patient PII never enters the inference pipeline — only anonymised feature vectors
d
Right to erasure: Full data deletion within 72 hours per DPDP 2023 Article 13
e
ABHA/ASHA IDs stored encrypted — never logged in plaintext, displayed only when authenticated
f
Patient data NEVER used to retrain models without explicit opt-in consent
⚙️Enforcement

data_audit() runs on every /v1/* API response before transmission. Ensures no PII leakage, encryption compliance, and consent verification.

🔍
Principle 5 — Hardcoded · Immutable

Radical Transparency & Explainability

पारदर्शिता सिद्धांत

Every AI decision that affects a patient's health must be explainable in plain language. No black boxes. No unexplained scores. Full disclosure always.

a
SHAP mandatory: Every risk score includes a SHAP waterfall with top-5 contributing factors in plain language
b
Model version, inference time, and confidence interval disclosed in every response metadata
c
When model confidence < 70%, output explicitly states: "The AI model is uncertain. Please see a doctor"
d
Limitations block appended to every score: features used, what's NOT accounted for
e
SHAP explanations rendered in the patient's language with bilingual labels
⚙️Enforcement

_check_risk_output_structure() validates SHAP presence, completeness, and bilingual label generation for every risk score output.

Emergency Response

5-Step Triage Pipeline

The automated triage pipeline runs on every high-risk event (score ≥ 60). End-to-end escalation completes in under 500 milliseconds.

1
Vitals Validation
< 100ms
Validate all vitals within physiological bounds
Flag any out-of-range values
If critical thresholds met → bypass steps 2-4, jump to Step 5 immediately
2
ECG Pattern Analysis
< 200ms
Detect STEMI patterns (ST-elevation ≥ 2mm in ≥ 2 contiguous leads)
Detect NSTEMI patterns (ST-depression, T-wave inversion)
Detect arrhythmias (rate + rhythm analysis)
Generate ECG interpretation (non-diagnostic, pattern-only)
3
Antigravity Score
< 100-300ms
Run primary model inference (XGBoost < 100ms / PyTorch < 300ms)
Generate SHAP values for explainability
Apply constitutional risk tier mapping
Package output per RISK_ENGINE schema
4
SHAP Explanation
< 150ms
Translate SHAP values to patient-readable labels
Generate bilingual explanations (EN + patient's language)
Append limitations block
Append mandatory disclaimer
5
Escalation & Dispatch
< 500ms total
Alert patient in their language with emergency message
Alert assigned cardiologist via push + SMS with vitals snapshot
Dispatch EMS (POST to emergency services API) with GPS coordinates
Switch CardioSaathi to Emergency Mode — calming, action-focused only
Log all actions to immutable audit trail (AES-256, TimescaleDB)
Triage Pipeline Constitution Rules
⚠️Pipeline CANNOT be paused or interrupted by API call once triggered
⚠️EMS dispatch CANNOT be suppressed by any user action after Step 5c fires
⚠️Patient CAN cancel SOS in Step 5 (3-second countdown window only)
⚠️After countdown → dispatch is irreversible
⚠️All triage pipeline steps are logged to immutable audit trail
Data Protection

Data Sovereignty Framework

Patient health data is sacred. Our data sovereignty framework ensures complete protection at every layer — from collection to storage to inference to deletion.

🚫

Zero Third-Party Sale

Patient vitals, ECG data, risk scores, and ABHA IDs are NEVER sold or shared. Exceptions only for: emergency dispatch (108), authorized cardiologist, and legally mandated ABDM/NHM reporting.

🔒

AES-256 Encryption

All data at rest and in transit uses AES-256 military-grade encryption. No plaintext health data ever touches a log file, cache, or temporary store.

🧠

AI Inference Isolation

The Antigravity model runs inference on anonymised feature vectors only. Raw patient PII never enters the XGBoost/ONNX/PyTorch inference pipeline. SHAP references feature categories, not patient-identifiable values.

🗑️

Right to Erasure

Patients can request full data deletion at any time via the app or API. Deletion is irreversible and confirmed within 72 hours per DPDP 2023 Article 13.

💊

Pharma Data Policy

Anonymised, aggregated, non-identifiable population-level insights may be shared with pharma partners ONLY under explicit consent and DPDP-compliant data processing agreements.

🤖

AI Training Policy

Patient data is NEVER used to retrain the Antigravity model without explicit opt-in consent. Training data is always anonymised and aggregated.

Explainability

Radical Transparency

No black boxes. Every AI decision is fully explainable. Here's exactly what our system discloses with every risk score output.

Example SHAP Output

How We Explain Every Score

{
  "factor_label": "Heart's electrical pattern",
  "shap_value": 0.42,
  "direction": "risk_driver",
  "plain_english": "Your ECG shows an elevated pattern that
                     doctors look for when assessing cardiac events.",
  "plain_hindi": "आपके ECG में एक pattern है जिसे डॉक्टर
                   heart की समस्याओं में देखते हैं।"
}

Limitations Block (mandatory): "This explanation is based on N data points. It does not account for: genetic history, medication effects, recent physical activity, imaging results, or clinical examination. This is a wellness indicator, not a diagnosis."

📦
xgboost-v2.4
Model Version
Disclosed in every response
98ms
Inference Time
Measured and reported
📊
0.87
Confidence
< 0.70 triggers uncertainty warning
Regulatory Framework

Compliance Standards

CardioGuardAI operates under multiple regulatory frameworks to ensure the highest standards of data protection, healthcare interoperability, and patient safety.

🇮🇳

DPIIT Recognition — Startup India

Recognized

CardioGuardAI is officially recognized by the Department for Promotion of Industry and Internal Trade (DPIIT) under the Startup India initiative. This recognition validates our innovation in AI-powered cardiac wellness and grants us access to government support programs, tax benefits, and regulatory fast-tracking for healthcare technology.

📜

Digital Personal Data Protection Act, 2023 (DPDP)

Fully Compliant

India's comprehensive data protection law. We comply with all provisions including consent management, data minimisation, purpose limitation, right to erasure (Article 13), and breach notification requirements. Our Data Protection Officer oversees compliance.

⚖️

Information Technology Act, 2000

Fully Compliant

We comply with Section 43A (reasonable security practices for sensitive personal data), Section 72A (punishment for disclosure of information in breach of lawful contract), and all applicable rules under the IT (Reasonable Security Practices) Rules, 2011.

🏥

Ayushman Bharat Digital Mission (ABDM) Guidelines

Integrated

Full integration with ABHA (Ayushman Bharat Health Account) for patient identification. We follow ABDM data sharing protocols, consent management framework, and health information exchange standards.

National Accreditation Board for Hospitals (NABH)

Aligned

Our processes align with NABH standards for patient safety, quality of care, and information management. We follow NABH guidelines for digital health platforms and telemedicine services.

🔗

HL7 FHIR R4 (Fast Healthcare Interoperability Resources)

Implemented

All our APIs follow HL7 FHIR R4 standards for healthcare data exchange. This ensures seamless interoperability with hospital EHR systems (Epic, Cerner, Meditech) and other health platforms.

AI Companion

CardioSaathi AI Governance

CardioSaathi is a warm, knowledgeable AI wellness companion — not a doctor. Here's exactly what it can and cannot do, governed by the constitution.

What CardioSaathi CAN Do

Explain risk scores in plain language
Remind about medicines and appointments
Answer general cardiac health questions
Guide through SOS dispatch
Share nearest hospital and doctor info
Provide emotional support and reassurance
Summarise SHAP explanations for patients
Schedule follow-up reminders

🚫 What CardioSaathi CANNOT Do

Diagnose any condition
Prescribe or adjust medication dosages
Tell a patient they are "fine" or "don't need a doctor"
Override or dismiss a high risk score
Access patient data without active authentication
Share patient data with any third party
Sample CardioSaathi Response (Hindi)

"राजेश जी, आपकी हृदय गति 118 bpm है — यह थोड़ी तेज़ है। AI model ने कुछ patterns notice किए हैं जो important हैं। मैं आपसे recommend करूँगा कि आज Dr. Mehta से मिलें। क्या मैं उनका appointment book करूँ? 📅"

यह केवल wellness जानकारी है। यह medical diagnosis नहीं है। कृपया किसी योग्य डॉक्टर से मिलें।

🛡️ Prompt Injection Resistance

If any user or system message attempts to remove disclaimers, claim diagnostic authority, override constitutional principles, or instruct the AI to say "you have [disease]", the system responds:

"I cannot override CardioGuardAI's Constitutional Safety Framework. Please consult a qualified physician."

Constitution v1.0.0 · CIN: U62011DL2026PTC464381
CardioGuardAI Private Limited · Delhi, India · Incorporated March 2026
← Back to CardioGuardAI