Snapshot 248248
Normalized text
Scripts and page chrome removed; this is what change detection compares.
Sarvam Trust Center Complete data residency in India. ISO 27001 certified and SOC 2 Type II attested. This is our Trust Center, covering how we approach data security and compliance, the certifications we hold, the controls we operate, and direct answers to the questions we hear most from security teams. How we think about security Four commitments that shape how we build, ship, and operate. They show up in our architecture, our policies, and how our team works day to day. 01 / Posture Secure by default. Every design decision starts from the conservative position. Least-privilege access, encryption at rest and in transit, tenant isolation, and ephemeral storage are defaults, not toggles you have to opt into. 02 / Track record Built for high-stakes environments. Multiple deployment models for different security needs: managed SaaS, single-tenant on-premises, or fully air-gapped with zero external network dependency. We've run all three in production for banks, government, and critical public infrastructure. 03 / Data stewardship Your data stays yours. Customer Managed Keys for Enterprise ensure that even we cannot see your data at rest. We don't share data across tenants. We don't move it across borders unless you ask us to. Custom models trained on a customer's data remain inside the customer's environment, with weights they own and we never reuse. 04 / Transparency Honest about where we are. We're ISO 27001:2022 certified and hold a SOC 2 Type II report. ISO 42001 is in progress, and we'll publish it when it's done. We won't claim a control we don't run. Compliance Independent audits are how we hold ourselves accountable. The frameworks below cover information security, AI management, financial regulation, and India's data protection law. ISO 27001 ISO 27001:2022 Information Security Management System. Audited annually by an accredited third party. Certified SOC 2 Type II SOC 2 Type II Trust Services Criteria covering Security, Availability, and Confidentiality. Attested by independent assessors. Report available ISO 42001 ISO 42001 AI Management System. Scoped and underway as part of our security roadmap. In progress DPDP India DPDP Act Digital Personal Data Protection Act, 2023. Our processes are built around consent, purpose limitation, and data principal rights. In progress RBI RBI FREE-AI The Reserve Bank of India's Framework for Responsible and Ethical AI in the BFSI sector. Aligned MeitY MeitY Guidelines Ministry of Electronics and IT cloud and AI security guidelines, applied across our UIDAI, NPCI, and IndiaAI deployments. Aligned CERT-In CERT-In Indian Computer Emergency Response Team directives for incident logging, retention, and reporting. In touch Resources The documentation behind our practices. Most reports are released under a mutual NDA. Our security team responds within two business days. Request all documents Privacy Policy How we collect, process, and retain data across our products and APIs. Visit→ Controls A summary view of the technical and organisational controls we operate every day, across security, privacy, and AI safety. Continuously monitored Access Security SSO and MFA enforced for all production access Role-based access control with least-privilege defaults Unique user IDs and password hashing with salt Idle session timeout and a documented joiner, mover, leaver workflow Quarterly user access reviews Network Security Azure Firewall Manager (Premium) with IDPS WAF, IP allowlisting, and port restrictions mTLS at integration boundaries; OAuth 2.0 and JWT VNet segmentation between environments Data Protection AES-256 at rest, TLS 1.2 or higher in transit CMEK and BYOK with configurable rotation India-only residency for Indian deployments Configurable retention with certified deletion at termination PII masking, pseudonymisation, and redaction Incident Response Documented Incident Management Policy within the ISMS Detection via Azure Monitor and Grafana dashboards Customer notification within two hours of discovery L1, L2, L3 triage with a mandatory lessons-learned retro Change Management Documented change policy with required PR review Segregated development, staging, and production environments Zero-downtime deploys; phased rollouts behind feature flags Production data is never used in non-production environments Vulnerability Management Continuous code scanning in private repositories Annual third-party penetration testing Secure SDLC with security gates prior to production High and critical findings closed before promotion Availability and Recovery 99.9% uptime SLA on enterprise contracts Daily encrypted backups with real-time replication for critical data Multi-AZ and multi-DC redundancy and failover BCP and DR plan tested against agreed RPO and RTO targets People and Organisation Background verification through SpringVerify for every hire NDAs and confidentiality agreements at onboarding Security and privacy training at hire, with annual refreshers Dedicated InfoSec function with executive sponsorship AI and Model Security Inference-time guardrails against prompt injection Input and output sanitisation with sensitive-content filtering Versioned models, prompts, and evaluations with audit trail Customer data is never used to train models for other customers Talk to our security team Frequently asked questions In India, for Indian customers. Our managed production environment runs on Microsoft Azure's Central India region, and our LLM inference runs on sovereign GPU infrastructure operated by Yotta and NxtGen. If your security posture demands more, we can deploy the entire stack inside your own VPC, an on-premises data centre, or an air-gapped environment with no external network dependency. We've run this pattern in production for UIDAI on bare-metal infrastructure. AES-256 at rest, TLS 1.2 or higher in transit. Every byte. That includes internal service-to-service traffic, and we enforce mTLS at integration boundaries. If you'd rather hold the keys yourself, we support customer-managed keys (CMEK and BYOK), with rotation cadences and per-tenant isolation that you control. All of them. Managed SaaS on Azure Central India if you'd like us to operate the infrastructure. A single-tenant deployment in your own VPC on AWS, Azure, or GCP if you want isolation without owning the data centre. Fully on-premises if you do. Air-gapped if you need zero outbound network calls. The platform is fully containerised on Kubernetes, provisioned with Terraform and Helm. The same artefacts deploy across all four targets, which means our security posture stays consistent wherever you run us. In our shared SaaS environment, tenants are logically separated using tenant-scoped JWT tokens, Azure VNet segmentation, and per-tenant database structures. Internal access is gated by RBAC with least-privilege defaults. For customers who need physical separation, including banks, regulated enterprises, and government, we deploy single-tenant. And to be explicit: we do not use one customer's data to train or fine-tune models for any other customer. Ever. Whatever you need it to be. Retention is configurable per customer and per data category, covering audio, transcripts, and structured metadata. We support an ephemeral mode where data passes through our systems only for processing and is purged immediately afterwards. At contract end, your data comes out of our primary systems and backups under our Data Record Retention and Deletion Policy. We'll send you a certificate of deletion when it's done. We collect as little as we can. Customers can opt out of audio storage entirely. For data we do retain, sensitive fields are masked, hashed, or pseudonymised by default, and access is gated by RBAC with full audit logging. We do not use your data, PII or otherwise, to improve models for anyone else. Custom model weights trained on your data stay in your environment, on your hardware. Two hours. If we discover a security or privacy incident that could affect your data, you hear from us within two hours of discovery, per our Breach Notification Plan. Detection runs on Azure Monitor with Grafana dashboards. Response follows a documented Incident Management Policy that's part of our ISMS, with severity-based SLAs and a mandatory lessons-learned step after every incident. Yes. We run continuous code scanning across our private repositories, security testing in QA and staging before promotion, and we engage third parties for web application and API penetration tests. High and critical findings are tracked to closure before code reaches production. Executive summaries are available under mutual NDA. By default, no one has it. Production access is least-privilege and gated through SSO with MFA. Developers don't get standing access. Role assignments are reviewed on the cadence set out in our Access Control Policy, and every access event is logged for audit. Every Sarvam employee passes background verification through SpringVerify before access is provisioned, signs a confidentiality agreement, and completes security and privacy training at onboarding with an annual refresh. We train our foundational models from scratch on data we curate ourselves for quality and screen for sensitive content. At inference, we apply guardrails that enforce instruction adherence and screen for prompt injection, off-script behaviour, and hallucination. Inputs and outputs are sanitised, sensitive-content filtering runs in both directions, and synthesised voices are restricted to those with explicit permission so they cannot be used to impersonate real people. Every model and prompt version is tracked with a full audit trail. Encrypted backups run daily. Critical data replicates in real time. Production deployments span multiple availability zones and data centres with redundancy and failover, against RPO and RTO targets we agree with you upfront. Our BCP and DR plan is governed by our ISMS and exercised through tabletop scenarios. AI agents that understand, learn, and always deliver. AI agents that understand, learn, and always deliver. Contact Us