🧠 AI & Governance
AI Risk & Governance Framework
Formal AI risk register, incident response protocol, model performance monitoring, sub-processor table, and versioned changelog — alongside our DPDPA 2023 compliance posture and human oversight design. Updated April 2026.
🇮🇳
India Data Residency
All data stored in AWS Mumbai. Nothing leaves India.
👤
Human-in-the-Loop
Zero auto-submission. CA certifies every QPR.
🔍
Explainable AI
Every extracted field cites its source document and line.
⚖️
DPDPA 2023
We are a Data Processor. Your org is the Fiduciary.
⚖️ DPDPA 2023 — Our role and yours
Digital Personal Data Protection Act 2023 · Rules notified November 2025
The most important thing to understand: Under DPDPA 2023, ReraDesk is a Data Processor — we process data on your documented instructions. Your organisation (developer, CA firm, or lender) is the Data Fiduciary — you determine the purpose and bear primary regulatory liability (up to ₹250 Crore). This distinction is non-shiftable under §10 DPDPA 2023.
🏗
What your organisation must do (Data Fiduciary)
- Determine why personal data is processed
- Obtain and manage buyer consents
- Respond to data principal rights requests
- Report breaches to the Data Protection Board
- Maintain a data map of all processing activities
YOUR responsibility
🛡
What ReraDesk does (Data Processor)
- Process data only per your instructions
- Maintain AES-256 encryption and access controls
- Notify you of breaches within 72 hours
- Delete data per your retention instructions
- Never use your data for our commercial purposes
ReraDesk responsibility
ReraDesk's DPDP Consent Manager (available in-app) provides tools to help your organisation manage buyer opt-ins, generate data maps, and draft breach notifications — but the legal obligation to act on these remains with your organisation as Data Fiduciary.
🧠 AI model design — how extraction works
Every AI suggestion is traceable, auditable, and requires human confirmation
Step 1 — Document ingestion
You upload bank statements, cost certificates, architect reports, or QPR drafts. Files are encrypted at rest immediately on receipt (AES-256). They do not leave India at any point in this pipeline.
AWS Mumbai · Encrypted in transit (TLS 1.3) and at rest (AES-256)
Step 2 — AI extraction (Amazon Bedrock, Mumbai)
Amazon Bedrock models extract QPR field values — financial progress, physical completion, unit sales, escrow balances. Bedrock is invoked in the ap-south-1 (Mumbai) region. Your documents are not used to train Bedrock's base models.
Amazon Bedrock AgentCore · Mumbai region · No model training on your data
Step 3 — Source traceability
Every extracted value is tagged with: the source document name, the specific page and line it was found on, a confidence score (High / Medium / Low), and the exact text snippet it was extracted from. Users can click any value in the QPR wizard to see its source.
Explainable AI · Full source citation per field · Confidence scoring
Step 4 — Human review (mandatory)
The QPR wizard requires the authorised user to review every extracted field. Low-confidence fields are flagged for explicit confirmation. No field proceeds to the submission step without user review.
Human-in-the-loop · Mandatory review · Cannot be bypassed
Step 5 — CA certification (gate)
A hard gate requires the Chartered Accountant to enter their full name, ICAI Membership Number, and firm name, then check a certification checkbox — affirming professional liability under the ICAI Code of Ethics and RERA 2016. The submission button is disabled until both the promoter declaration and CA certification are complete.
DSC certification required · Promoter declaration required · No auto-submit
Step 6 — Audit trail locked
On submission, a SHA-256 timestamped audit trail is created: which user filed, which CA certified, what fields were extracted vs manually entered, and what documents were cited. This trail is immutable and available for court proceedings or RERA authority inspection.
SHA-256 timestamp · Immutable · Court-admissible under §65B Evidence Act
🇮🇳 Data residency and infrastructure
All ReraDesk data remains in India at all times
| Component |
Provider |
Location |
Status |
| Application hosting |
Cloudflare Pages |
Global CDN (static assets only — no personal data) |
● Live |
| Database (project data, filings) |
Supabase (PostgreSQL) |
Mumbai (ap-south-1) · Row-Level Security enabled |
● Live |
| Document storage |
AWS S3 |
Mumbai (ap-south-1) · AES-256 SSE |
● Live |
| AI extraction |
Amazon Bedrock |
Mumbai (ap-south-1) · No cross-region data transfer |
● Live |
| Authentication |
Supabase Auth |
Mumbai · Passwords bcrypt-hashed |
● Live |
| Payment processing |
Razorpay |
India · PCI-DSS compliant · No card data stored by ReraDesk |
● Planned |
| Product analytics |
PostHog |
EU Cloud (anonymised, no personal data) · Opt-out available |
● Planned |
🔐 Access control and security
Who can see your data, and under what conditions
👥
Organisation-level isolation
Supabase Row-Level Security (RLS) ensures each organisation can only query its own data. No cross-tenant data access is possible at the database level — this is enforced by the database itself, not just application code.
🧑💻
ReraDesk staff access
No ReraDesk employee can access your project data without a documented, time-limited authorisation. All staff access is audit-logged. Access is limited to support engineers with a valid support ticket open.
🔑
DSC and private keys
ReraDesk never receives, stores, or transmits your DSC private key. Digital signature operations use your local browser extension or USB token. The private key never leaves your device.
⚠️
Breach response
In case of a confirmed breach, affected organisations are notified within 72 hours. Notification includes: what data was affected, likely impact, steps taken, and recommended actions for your organisation.
👤 Human oversight — the non-negotiable design principle
ReraDesk is a decision-support system, not an autonomous compliance engine
Our position on AI autonomy in compliance: RERA compliance has real legal and financial consequences for promoters and their CAs. ReraDesk is designed as a Human-in-the-Loop (HITL) system at every step that matters. AI extracts and suggests — humans review, certify, and decide. This is not a limitation; it is deliberate governance design.
Specific HITL checkpoints built into the platform:
- QPR wizard: Cannot proceed past Step 6 without promoter declaration checkbox + CA name, ICAI number, and certification checkbox
- GST-ITC reconciliation: IMS invoice Accept/Reject actions are manual — no auto-accept
- Form-7 Tally reconciler: CA certification button is disabled while any variance is unresolved
- 3-Account compliance score: Displayed as an analytical indicator with visible methodology — not a compliance certificate
- CIRP risk scores: Displayed with full methodology disclosure — not a legal insolvency opinion
- QR Embedder: Generates compliant QR codes — but uploading to promotional material is done by the user, not automated
The phrase "ReraDesk suggests · CA decides · RERA authority rules" is displayed prominently in the CA Co-Pilot dashboard as a permanent reminder of this hierarchy.
📊 AI model validation and accuracy
How we measure and report extraction accuracy
ReraDesk uses Amazon Bedrock foundation models for document extraction. Our approach to model governance:
- Confidence scoring: Every extracted field is tagged High / Medium / Low confidence based on model certainty. Low-confidence fields are flagged for mandatory user review.
- Source citation: Every extraction cites its source document, page, and line — enabling human auditors to verify independently
- No black-box outputs: ReraDesk does not display scores or recommendations without showing the underlying data and methodology
- Accuracy tracking: We track extraction accuracy across pilot filings. Current benchmark: 99%+ field-level accuracy on structured documents (banker's certificates, architect completion certificates). Accuracy is lower on unstructured narrative documents — these are flagged for manual review.
- Bias monitoring: We monitor for systematic extraction errors across states, document types, and languages. Known limitations: Hindi-language documents currently extract at lower accuracy — flagged for review automatically.
- No training on your data: Amazon Bedrock's model usage agreement prohibits use of API inputs for model training. Your documents do not improve the base models.
Accuracy disclaimer: "99%+ submission success in beta" refers to tracked outcomes across pilot filings where extraction was reviewed and certified by a CA. Individual results vary by document quality, RERA portal uptime, and state-specific requirements. This is not a guarantee of acceptance by any RERA authority.
🏛 Regulatory language — what we say and what we don't
Precise framing to avoid misleading claims
| We say | We don't say | Why |
|---|
| "Built for MahaRERA Order 46C compliance" |
"MahaRERA certified" or "MahaRERA approved" |
We have not received official certification from MahaRERA |
| "Designed for TNRERA 100/70/30 mandate" |
"TNRERA authorised tool" |
We are a private platform, not a government-authorised system |
| "CREDAI BuildSmart innovation applicant" |
"CREDAI certified" or "CREDAI endorsed" |
CREDAI is a trust mark; it is not a regulatory clearance |
| "Analytical decision-support tool" |
"Compliance guarantee" or "Zero penalty guarantee" |
No tool can guarantee regulatory outcomes |
| "99%+ accuracy benchmarked in beta" |
"100% success rate" |
ASCI guidelines require qualified performance claims |
🚨 AI Risk Register — Identified risks and mitigations
We maintain a formal AI risk register, reviewed quarterly. The table below reflects risks identified as of April 2026 and the controls in place.
| Risk | Likelihood | Impact | Control | Residual risk |
|---|
| Extraction hallucination — AI assigns incorrect value to a QPR field |
Medium |
High |
Source citation mandatory for every field; low-confidence fields blocked from submission until human review; CA certification gate |
Low |
| Confidence score miscalibration — High confidence assigned to incorrect extraction |
Medium |
Medium |
Monthly accuracy audits against known-correct QPR datasets; model recalibration if F1 score drops below 0.95 |
Medium |
| Model drift — Bank statement / document formats change over time |
Low |
Medium |
Weekly extraction drift monitoring; automated alert if field-level accuracy drops >3% week-on-week |
Low |
| Prompt injection via malicious documents — Uploaded document attempts to alter AI behaviour |
Medium |
High |
Document processing in isolated sandbox (Bedrock); extracted text never fed back as system prompt; output sanitised before display |
Low |
| Regulatory language misuse — Users misrepresent ReraDesk outputs as official compliance certification |
Medium |
Medium |
Disclaimer on every output screen and generated document; governance page; terms of service §4; ASCI-compliant marketing claims |
Low |
| CA certification gate bypass — Technical exploit allows submission without CA sign-off |
Low |
Critical |
Server-side validation of certification flag independent of client-side UI; audit log records certification event; submission without log entry triggers alert |
Low |
| Data breach — escrow / financial data — Unauthorised access to stored QPR data |
Low |
Critical |
AES-256 at rest, TLS 1.3 in transit, RLS per-organisation isolation, bcrypt auth, MFA available, 72-hour breach notification per DPDPA 2023 §8(6) |
Low |
Risk register last reviewed: April 2026. Next scheduled review: July 2026. Responsible: Data Protection Officer — [email protected]
🔥 Incident Response Protocol
ReraDesk follows a documented incident response process for AI-related failures and data security events. The process is aligned with DPDPA 2023 §8(6) and CERT-In guidelines.
Severity 1 — Critical
Data breach · CA gate bypass · Mass extraction failure
Response time: 1 hour. DPDPA breach notification within 72 hours. Service suspended pending investigation. All affected users notified by email.
Severity 2 — High
Extraction accuracy drop · API outage · Auth failure
Response time: 4 hours. Affected feature degraded or suspended. Users notified in-app. Root cause published within 7 days.
Severity 3 — Medium
Performance degradation · Single-field extraction error
Response time: 24 hours. Monitored and patched in next release. Affected users notified if data integrity impacted.
Severity 4 — Low
UI glitch · Non-critical feature failure
Response time: 72 hours. Fixed in scheduled release. Logged in public changelog.
Reporting a security or AI incident
Email [email protected] with subject line "INCIDENT — [Severity] — [Brief description]". We acknowledge within 2 hours during business hours (09:00–18:00 IST). For critical incidents outside business hours, include "URGENT" in the subject.
📈 Model Performance Monitoring
AI extraction quality is monitored continuously. We measure performance at three levels:
| Metric | Target | Current (Beta) | Alert threshold | Action if breached |
|---|
| Field-level extraction accuracy |
≥ 97% |
99.1% |
Below 95% |
Automatic model rollback to previous version; engineering review within 24h |
| Confidence score calibration |
F1 ≥ 0.95 |
0.97 |
F1 below 0.92 |
Recalibration sprint; additional training data review |
| Low-confidence field rate |
< 8% of fields |
5.3% |
Above 15% |
Document format analysis; prompt engineering review |
| CA override rate |
< 5% of extractions |
3.1% |
Above 10% |
Root cause analysis on overridden fields; model fine-tuning |
| False high-confidence rate |
< 1% |
0.4% |
Above 2% |
Immediate confidence threshold recalibration; user notification |
Performance data is collected anonymously (no PII, no document content). Monthly performance reports are reviewed by the AI governance committee. Significant changes are disclosed in the changelog below.
Model versioning policy
All model updates are versioned (e.g., v2.1.4). Breaking changes (accuracy threshold changes, new field types, confidence recalibration) are announced via in-app notification 7 days before deployment. Emergency rollbacks are executed without notice but disclosed within 24 hours.
🔗 Third-Party Sub-Processors
Under DPDPA 2023 §10, we are responsible for the acts of our sub-processors. The table below lists every third-party service that processes personal data on our behalf.
| Sub-processor | Purpose | Data processed | Location | DPDPA basis |
|---|
Amazon Web Services
AWS S3, Lambda, Bedrock |
Document storage, AI extraction, serverless compute |
Uploaded QPR documents, extracted field data |
ap-south-1 (Mumbai) |
Data Processing Agreement in place; AWS Mumbai DPDPA-aligned |
Supabase
PostgreSQL, Auth, Storage |
Database, authentication, row-level security |
User accounts, QPR metadata, audit logs |
ap-south-1 (Mumbai) |
DPA in place; data does not leave India region |
Cloudflare
CDN, WAF, Pages |
Web delivery, DDoS protection, edge caching |
IP address, request metadata (no body content) |
Global edge (no personal data stored) |
No personal data persisted; edge metadata deleted within 24h |
Razorpay
Payment processing |
Subscription billing, payment collection |
Name, email, payment instrument (not stored by us) |
India (RBI-regulated) |
PCI DSS compliant; payment data never touches ReraDesk servers |
PostHog
Product analytics |
Usage analytics, funnel tracking |
Anonymised event data, user ID (no PII unless identified) |
US Cloud (posthog.com) |
No regulatory personal data; only product usage events; opt-out available |
Formspree
Beta application forms |
CA beta waitlist form submission |
Name, email, firm name from CA beta form only |
US (formspree.io) |
Consent collected at form submission; data deleted after onboarding |
Crisp
Customer support chat |
In-app support and onboarding chat |
Name, email, chat conversation content |
EU (crisp.chat) |
Chat data retained for support quality only; deleted on request |
Sub-processor list last updated: April 2026. Changes to sub-processors that affect personal data will be notified to enterprise subscribers at least 30 days in advance. To request the current DPA for any sub-processor, email
[email protected].
📋 Change Management & Governance Changelog
Material changes to AI behaviour, data processing, security controls, or regulatory compliance are documented here. This log is the authoritative record of platform governance evolution.
v1.4 — April 2026
Current version
- Added UDIN 5-parameter validation (MRN + Form No + AY + PAN + UDIN) — ICAI mandate April 1, 2026
- Added GSTIN active/suspended status validation via GST API integration
- Added MahaRERA Grade Simulator — 10-criterion framework per Grading Circular 2024
- Added Safari Retreats SC ruling (2024) guidance in ITC Recovery tooling
- Razorpay payment integration — UPI AutoPay + GST invoice generation
- Crisp chat widget deployed — user identification wired to auth system
- AI risk register formalised — 7 identified risks, controls documented
- Sub-processor table published — 7 processors listed with DPDPA basis
v1.3 — March 2026
- QPR low-confidence field guard — blocks submission until all flagged fields explicitly reviewed
- Overdue → Compliant status transition wired to QPR filing event
- GST-ITC IMS 2026 tab — Invoice Management System Accept/Reject queue
- Form-7 Tally Reconciler — Tally XML import, variance detection, CA certification gate
- CREDAI BuildSmart application submitted — Theme 03 (RERA Single-Window)
- SEO canonical + Open Graph tags added to all pages
- Cross-URL UTM funnel tracking deployed (index → ca → app)
- DPDPA 2023 Rules (notified November 2025) incorporated into Privacy Policy
v1.2 — February 2026
- MahaRERA QR Embedder — Order 46C compliance, fine shield calculator
- Sweeper (100/70/30 compliance monitor) — live score with visible methodology
- CIRP Risk Dashboard — escrow gap analysis, project-wise insolvency (IBC 2025)
- CA Co-Pilot multi-tenant dashboard launched — UDIN tracker, PIE self-review notice
- PostHog and GA4 analytics deployed — user identification, QPR event tracking
- Privacy, Terms, and Governance pages published (DPDPA 2023 compliant)
- Supabase Auth URLs configured — live signup and email confirmation active
v1.0 — January 2026
Initial launch
- QPR Filing Wizard v1 — 7-step workflow, AI extraction, CA certification gate
- Source tracing architecture — every field tagged to source document, page, line
- SHA-256 audit trail — immutable, §65B IT Act 2000 admissible
- Zero auto-submit policy established — human confirmation required at all submission points
- AWS Mumbai data residency enforced for all project data
- ASCI compliance review completed — all marketing claims qualified
- Initial deployment on Cloudflare Pages — reradesk.in live
Governance changelog is updated with every material platform release. Enterprise subscribers receive change notifications by email. To subscribe to change notifications:
[email protected] with subject "Subscribe to governance updates."