सिंथेटिक डेटा — केवल प्रदर्शन हेतु / SYNTHETIC DATA — DEMONSTRATION ONLY · CRISP AI Portfolio Platform
CA

VISHWAS

विश्वासRFP §6Responsible AI & Compliance

VISHWAS — Responsible AI & Compliance Console (RFP §6)

Every AI capability in the portfolio is governed by VISHWAS: fairness monitoring, transparency (Model Cards), explainability, accountability (audit logs), human oversight, DPDP/ISO 27001/CERT-In/OWASP compliance, and prompt-injection guardrails.

16/16
Compliance Flags On
6
Security Controls
12
Model Cards Published
20+
Audit Entries (recent)

Responsible AI

6/6 enabled

Accountability

ACCOUNTABILITY

Immutable audit log; every AI output traceable to prompt version, model, provider, sources, user

Bias Monitoring

BIAS_MONITOR

Scheduled fairness metric recompute; drift detection; alerting

Explainability

EXPLAINABILITY

Every AI output carries "Why?" button — citations, SHAP, Grad-CAM, attribution weights

Fairness Monitoring

FAIRNESS

Disparity dashboard — outcome/score distributions by gender, social category, district. Flag disparate impact ratio < 0.8

Human Oversight

HUMAN_OVERSIGHT

Human-in-loop queues with override rate metric; confidence-threshold routing

Transparency — Model Cards

TRANSPARENCY

Model Card per AI feature: purpose, training data, limitations, operating point, TRL

Security

6/6 enabled

CERT-In Guidelines

CERTIN

180-day log retention, incident response runbook, NTP sync, vulnerability disclosure

DPDP Act 2023

DPDP

Consent registry, purpose limitation, retention policy, data-principal rights, data-residency enforcement

Encryption Standards

ENCRYPTION

TLS 1.3 in transit; AES-256-GCM at rest; envelope encryption for API keys

ISO 27001

ISO27001

Information security management system controls

OWASP Top 10

OWASP

Web application security controls + pentest checklist

OWASP LLM Top 10

OWASP_LLM

Prompt injection defence, output encoding, model DoS limits, sensitive-info filters

Governance

4/4 enabled

Audit Logs

AUDIT_LOGS

Tamper-evident hash-chained audit trail with search/export UI

Data Retention Policies

DATA_RETENTION

Per-entity retention policy; automated purge; retention report

Model Governance

MODEL_GOV

Model registry: version, approver, deployment date, eval scores, rollback

User Access Control

RBAC

RBAC + ABAC (posting-level scoping: Block officer cannot query state data)

Model Cards — Transparency (RFP §6.2)
12 published
Every AI feature has a published Model Card documenting purpose, training data, limitations, operating point and TRL.
AAROGYA/cxr-triage
GLM-4V Vision vv1.0
TRL 6

Chest X-ray abnormality triage

Data: Public CXR datasets (NIH ChestX-ray14, CheXpert)
Limits: Screening only — not diagnosis; TB endemic bias
Op: 95% sensitivity, 68% specificity (triage-tuned)
ABHILEKH/devanagari-ocr
GLM-4V Vision vv1.1
TRL 7

Hindi+English OCR for govt documents

Data: Scanned Khasra/Khatauni, certificates, letters
Limits: Faded/handwritten annotations lower accuracy
Op: field confidence < 0.8 → human review
BHU/change-detect
GLM-4V Vision vv1.0
TRL 6

Satellite change detection

Data: Sentinel-2 bi-temporal pairs, MP districts
Limits: Cloud cover, seasonal crop changes cause false positives
Op: NDVI/NDBI differencing + vision verification
GYAN/domain-llm
Fine-tuned GLM-4 vv1.0
TRL 7

Domain-specific LLM (Health/Agri/Revenue/Legal)

Data: MP govt domain corpora per department
Limits: General knowledge may be less accurate than base model
Op: temperature 0.3, domain-scoped responses
GYAN/private-gpt
On-Prem Llama-3.1-8B vv1.0
TRL 7

Private GPT on-premise deployment

Data: Uploaded govt documents only
Limits: No internet access; limited to indexed corpus
Op: RAG with citation-required, air-gap ready
KISAN/disease-detect
GLM-4V Vision vv1.0
TRL 6

Plant disease detection from leaf images

Data: PlantVillage + MP crop images (soybean, wheat, gram, cotton, paddy)
Limits: Limited to 12 MP-relevant diseases; field conditions vary
Op: confidence threshold 0.85 for auto-advice
NIRNAY/forecast
Statistical (Prophet-like) vv1.0
TRL 7

Scheme progress forecasting

Data: 5-year scheme progress time series
Limits: Assumes linear policy environment
Op: prediction interval, not point estimate
SAHAYAK/grievance-classify
GLM-4 LLM vv1.3
TRL 7

Grievance routing & urgency scoring

Data: Synthetic MP grievances (~50k)
Limits: Long-tail categories may misroute
Op: urgency score 0-100 with feature attribution
SAMVAAD/policy-qa
GLM-4 LLM vv1.2
TRL 6

Government knowledge assistant with RAG

Data: GoMP circulars, schemes, acts (~500 docs)
Limits: Cannot answer outside indexed corpus; Hindi dialects may vary
Op: temperature 0.2, citation-required, abstention < 0.6
SHIKSHA/adaptive
Bayesian Knowledge Tracing vv1.0
TRL 6

Adaptive learning path

Data: Curriculum knowledge graph (MP Board + NCERT)
Limits: Cold-start for new students
Op: target 70% success probability
SURAKSHA/fraud-detect
GLM-4 LLM vv1.0
TRL 6

Fraud detection for scheme disbursements

Data: Synthetic MP beneficiary + transaction data
Limits: Sophisticated collusive fraud may evade detection
Op: riskScore 0-100 with pattern attribution
SURAKSHA/threat-classify
GLM-4 LLM vv1.0
TRL 6

Security event threat classification

Data: NIST/CERT-In threat taxonomy + synthetic MP SOC events
Limits: Long-tail zero-day patterns may misclassify
Op: confidence threshold 0.8 for auto-response
Feature Flags (L3)
13/13 on
Human-in-the-loop Queue
AAROGYA.HUMAN_LOOP · AAROGYA
OCR Human Review Queue
ABHILEKH.HUMAN_REVIEW · ABHILEKH
Air-Gap Deployment Mode
GYAN.AIR_GAP · GYAN
Private GPT (On-Premise)
GYAN.PRIVATE_GPT · GYAN
Role-Based Access Control
GYAN.RBAC · GYAN
Grad-CAM Explainability
KISAN.GRADCAM · KISAN
DSS Live Sliders
NIRNAY.MCDA_SLIDERS · NIRNAY
Voice Bot (IVR)
SAHAYAK.VOICEBOT · SAHAYAK
WhatsApp Assistant
SAHAYAK.WHATSAPP · SAHAYAK
Fraud Detection Engine
SURAKSHA.FRAUD_DETECT · SURAKSHA
SOC Automation Playbooks
SURAKSHA.SOC_AUTOMATION · SURAKSHA
Fairness Dashboard
VISHWAS.FAIRNESS_DASH · VISHWAS
Prompt Injection Guardrails
VISHWAS.GUARDRAILS · VISHWAS
Audit Log — Recent 20
20
TimeActionModuleEntity
8/26/2026, 3:35:38 PMDISABLEGYANEnterpriseAiConfig/cmt9x2ru
8/26/2026, 3:35:26 PMENABLEGYANEnterpriseAiConfig/cmt9x2ru
8/26/2026, 3:06:12 PMCONFIG.UPDATE_PROVIDERCONFIGApiProvider/cmt9ouco
8/26/2026, 3:05:44 PMCONFIG.UPDATE_PROVIDERCONFIGApiProvider/cmt9ouco
8/26/2026, 1:04:17 PMVISHWAS.TOGGLE_FLAGVISHWASFeatureFlag/cmt9oucp
8/26/2026, 1:04:16 PMVISHWAS.TOGGLE_COMPLIANCEVISHWASComplianceFlag/cmt9oucp
8/26/2026, 1:04:05 PMCONFIG.DEMO_PRESETDEMO_DIRECTORApiProvider/live
8/26/2026, 1:04:05 PMCONFIG.UPDATE_PROMPTCONFIGPromptTemplate/cmt9ouco
8/26/2026, 1:04:04 PMCONFIG.UPDATE_PROVIDERCONFIGApiProvider/cmt9ouco
8/26/2026, 1:00:59 PMCONFIG.DEMO_PRESETDEMO_DIRECTORApiProvider/hybrid
8/26/2026, 11:59:30 AMAI.GENERATEBHU-
8/26/2026, 10:59:30 AMAI.GENERATEKISAN-
8/26/2026, 8:59:30 AMAI.GENERATESAHAYAK-
8/26/2026, 6:59:30 AMAI.GENERATESAMVAAD-
8/26/2026, 1:06:30 AMVISHWAS.ENABLE_FLAGVISHWASComplianceFlag/OWASP_LL
8/26/2026, 1:05:30 AMVISHWAS.ENABLE_FLAGVISHWASComplianceFlag/DPDP
8/26/2026, 1:04:30 AMVISHWAS.ENABLE_FLAGVISHWASComplianceFlag/ISO27001
8/26/2026, 1:02:30 AMCONFIG.SEED_PROMPTSCONFIG-
8/26/2026, 1:01:30 AMCONFIG.SEED_PROVIDERSCONFIG-
8/26/2026, 12:59:30 AMPLATFORM.BOOTCONFIG-

Fairness Dashboard (Concept)

Tracks outcome distributions by gender, social category, and district across every AI feature. Flags disparate impact ratio < 0.8 for review. In production, this card would render live charts from a fairness-metrics store; in this demo it visualises the methodology and wiring — every model card exposes its fairness operating point.

Gender parity
0.94
Social category
0.88
District parity
0.79
Override rate
12%

Prompt Injection Guardrail Demo (OWASP LLM §1)

The VISHWAS guardrail layer runs server-side before the LLM is invoked, checking every input against known prompt-injection patterns. Try a safe query first (it passes), then try an injection attempt (it gets blocked) — this is the OWASP LLM Top 10 §1 defence working live.

Quick try: