Our trust pages give the executive summary. This is the long form — the document we hand to security teams, auditors, and procurement. Everything below is what we actually ship in every deployment, and what every third party that touches your data does with it.
Two principles run through all of it. First, controls are architectural, not bolt-ons — PII redaction, audit logging, and tenant isolation live in the pipeline, not in a policy document. Second, India is the baseline and global certifications follow our enterprise expansion. We tell you what is live, what is in-flight, and what is on the roadmap, with dates.
Regulatory alignment.
Where we stand today across every standard relevant to our products and markets — the scope it covers, the products it touches, and its current status.
| Standard | Scope | Relevant to | Status |
|---|---|---|---|
| DPDP Act 2023 | India | All Indian deployments | Live — consent capture, retention classes, withdrawal, PII redaction pre-LLM; India residency by default |
| RBI Fair Practices Code | India BFSI | Lens — Helm (lending) | Live — interest disclosures, rejection reasoning, grievance routing automated at the agent layer |
| SEBI AI Guidelines | India Capital Markets | Future securities products | Aligned — Blueprint architecture designed to meet explainability and audit requirements |
| CDSCO MDR 2017 | India Medical Devices | Lens — KidneyCare | TRL 4→6 — on the CDSCO device pathway; target regulatory submission 2026–27 |
| ISO 27001:2022 | International | All enterprise clients | In-flight — ISMS underway; target certification Q4 2026; gap analysis complete |
| SOC 2 Type II | US Enterprise | US-headquartered clients | Roadmap — planned 2027 with US market entry; controls architecture already aligned |
| GDPR | European Union | EU data subjects / EU clients | Roadmap — architecturally consistent with DPDP; formal programme for 2027 EU expansion |
| HIPAA | US Healthcare | Lens — KidneyCare (US) | Roadmap — PHI controls, BAA capability, US-resident tier planned 2027 with FDA 510(k) |
Compliance by geography.
Where your data lives and which regulations govern it. India is our primary market today; the EU and US are on the roadmap with dated programmes.
India — primary market
All current deployments. Data residency in ap-south-1 (Mumbai) by default. No cross-border transfer without explicit consent. DPDP Act 2023 (Live), RBI Fair Practices (Live), SEBI AI Guidelines (Aligned).
European Union — roadmap
Target market from 2027. EU-resident storage tier to be established on AWS eu-central-1 (Frankfurt). Formal GDPR programme to follow, alongside EU AI Act risk tiers (both 2027).
United States — roadmap
US market entry planned for 2027, led by KidneyCare's FDA pathway. SOC 2 Type II and HIPAA BAA capability required for enterprise procurement (both 2027); FDA 510(k) for KidneyCare (2027–28).
Technical controls — every deployment.
These controls are architectural, not bolt-ons. The data flow that runs in every deployment, before any model is invoked: inbound data (WhatsApp · API · IoT · form · document) → PII redaction layer (Aadhaar · PAN · mobile · account no. stripped before any model call) → agent pipeline (T1 self-hosted OSS · T2 mid-cloud · T3 frontier · HITL checkpoints).
All model calls are logged. Every external API call, model invocation, and human decision is written to a WORM audit store with retention class enforced at the bucket policy level. Customer-managed keys (CMK) mean we never see plaintext at rest. Per-tenant key isolation is enforced at the storage layer.
- Immutable audit trail. Every external call, model invocation, and human decision written to a WORM store. Retention class enforced at bucket policy. Tamper-evident by design.
- Per-tenant customer-managed keys (CMK). Encryption at rest using CMK via AWS KMS. We never see your plaintext. Key rotation customer-controlled.
- PII redaction pre-LLM. Aadhaar, PAN, mobile numbers, and account numbers stripped before any model call. Redaction happens in the ingestion pipeline, not as an afterthought.
- India data residency by default. All deployments run in ap-south-1 (Mumbai). Private endpoints. No public egress. Cross-border transfer only with explicit customer consent and documented purpose.
- RBAC + tenant isolation. Row and column ACLs. Per-tenant CMK keys. Every query scoped to the user's permission set. No cross-tenant data leakage possible at the storage layer.
- Limbic veto — pre-execution safety stop. Unsafe action paths are halted before execution. Bedrock Guardrails for content safety on all frontier-tier calls. HITL routing for high-consequence decisions.
- Structured outputs + per-stage evaluators. Every agent stage produces typed, validated outputs. Per-stage evaluators catch hallucinations and drift before they propagate downstream.
Sub-processors — in full.
Every third party that touches your data. The list is reviewed quarterly, and customers are notified of any addition or substitution with 30 days notice.
| Vendor | Purpose | Data location | Data class | Vol. |
|---|---|---|---|---|
| AWS | Infrastructure · Bedrock LLM · KMS · S3 WORM | ap-south-1 (Mumbai) | all classes · encrypted (CMK) | 100% |
| Self-hosted OSS · tier 1 | LLM tier 1 — majority of agent calls | ap-south-1 (Mumbai) | in-tenancy · no egress | ~80% |
| Self-hosted OSS · tier 2 | LLM tier 2 — multilingual and document tasks | ap-south-1 (Mumbai) | in-tenancy · no egress | ~15% |
| Anthropic (via AWS Bedrock) | LLM tier 3 — complex reasoning and HITL explanation | US (when used) — PII-stripped before call | PII-stripped · no Aadhaar/PAN/acct no. | ~5% |
Self-hosted OSS models run entirely within the customer's AWS tenancy. No data is transmitted to third-party model APIs for ~95% of calls. For the 5% of calls routed to Anthropic via Bedrock, PII is stripped at the ingestion layer before the call is made.
DPDP specifics.
When you deploy an Aminobots system, your customer data is your data. We process it on your behalf, in your tenancy, under your encryption keys, and we never see your customer's personal information in plaintext. We retain processed records only as long as DPDP requires by class.
If your customer requests their data be deleted, the system supports that. If your customer requests their consent be withdrawn, the system supports that. If your auditor requests evidence of every data event, the immutable audit log produces it.
Retention by class
What we keep, for how long, and the legal basis for each.
| Class | Retention | Basis |
|---|---|---|
| KYC records | 7 years | RBI Master Directions |
| Customer interaction logs | 5 years | DPDP general retention |
| Transaction records | 10 years | RBI Banking Regulation Act |
| Clinical records (KidneyCare) | 5 years | ICMR 2017 guidelines |
| AI inference traces (non-PII) | 2 years | Internal audit + model improvement |
Consent, withdrawal, and the DPO
How your customers exercise their rights, and who to contact for a DPDP grievance.
- Consent capture & withdrawal. Consent is captured at ingestion. Every deployment supports consent withdrawal — the system marks the user's records and prevents further processing within 24 hours.
- Data deletion. On request, records are deleted within the legal grace period (typically 30 days). Audit log retention follows the retention-class table above.
- DPO contact. For DPDP grievances, contact our Data Protection Officer at [email protected]. Response within 5 business days; resolution within 30.
Responsible AI.
How we make agents safe to deploy. Three commitments hold across every system.
- Explainability — every decision is traceable. Every Helm credit decision shows the credit officer the underwriter agent's reasoning, the gate that triggered HITL, the confidence interval, and the policy clause invoked — a deposition-ready trail, not a black box.
- Human-in-the-loop — agents know when to stop. HITL checkpoints are built into the Blueprint, not added later. Credit decisions, clinical flags, and financial disbursements always have a defined human escalation path.
- Bias & drift — evaluated at every stage. Per-stage evaluators catch hallucination and distributional drift before it propagates. The CLEAR baseline is used to measure capability degradation over time.
For reviewers.
This post is the public reference. Under NDA, enterprise prospects can request the full security pack — architecture diagrams, sub-processor DPAs, pentest summary, ISMS scope statement, and the DPDP compliance questionnaire. The executive summary lives on our Trust & security page.