WinGrants AI
Security & data sovereignty summary.
A 2-page brief for IT, security, and procurement reviewers. Covers our certification posture, jurisdiction, AI routing controls, and the contractual commitments that travel with every customer engagement. The full Statement of Applicability and audit reports are available under a short mutual NDA — contact aj@wingrants.ai.
§1 Provider & legal
Legal entityHealthdev OÜ (trading as WinGrants AI)
JurisdictionRepublic of Estonia · EU member state
Registry code14588534
Registered officeMännimäe/1, Pudisoo küla, Kuusalu vald, Harju maakond, 74626, Estonia
Data protection contactaj@wingrants.ai
Governing lawEstonia · EU consumer-protection carve-outs preserved
§2 Certification & audit
StandardISO/IEC 27001:2022 — Information Security Management System
Certificate №GCAI-ISMS-LP84
Issued byGCAI Certification Services LLP (IAS-accredited)
First issued · Valid until23 Mar 2026 · 22 Mar 2029
Audit cycleAnnual external surveillance audits · full recertification every 3 years · documented internal audits in between. Findings flow into a corrective-action register reviewed by management.
ScopeWinGrants AI platform · development pipeline · internal corporate IT · EU data-centre footprint · support operations.
§3 Data residency & hosting
- EU-hosted by default. Multi-tenant deployment runs on AWS Frankfurt (eu-central-1). No transatlantic data flow for customer content under normal operation.
- Air-gap capable on Enterprise. Full platform shipped as a self-contained Docker Compose bundle with open-source model weights included. Zero outbound network calls required. Customer chooses where it runs — on-prem, sovereign EU cloud, or fully offline.
- Sub-processors. Minimised list, EU-resident where possible, published in the DPA. Annual reassessment under ISO A.5.19–5.23.
- Backups. Immutable WORM backups, 90-day retention, geographically distributed across EU. RPO 1h · RTO 4h. Quarterly DR tabletop exercises.
§4 AI routing & model controls
Zero data transfer to AI training pipelines. Contractually enforced in the DPA, technically enforced at the inference layer (training disabled on every endpoint we use, zero provider retention), and verifiable through tenant-tagged audit logs. Customer Data is never used to train or fine-tune any model — ours, an upstream model, or a third-party model.
- Open-source models only. Platform routes between Llama 4 Maverick, Qwen 3.6, Mistral Large 3, DeepSeek V4, Gemma 4, and gpt-oss-120B depending on task. No commercial closed-model API calls in the chain by default.
- Version pinning. Models pinned to specific versions; upgrades require written customer approval on Enterprise. The old model stays available in your tenancy until you choose to retire it.
- EU-resident inference. Inference endpoints in EU data centres by default. Regional pinning available on Enterprise.
- BYOK / HYOK. Bring-your-own-keys for Anthropic, OpenAI, Azure OpenAI, and EU-hosted providers on AI Enablement and Enterprise. HYOK (hold-your-own-keys) reserved for air-gapped Enterprise deployments.
§5 Identity, access, encryption
Encryption at rest
AES-256-GCM. Envelope encryption with per-tenant Data Encryption Keys protected by a KEK held in a hardware-backed key store (FIPS 140-2 L3 equivalent, EU-located).
Encryption in transit
TLS 1.3. CNSA cipher suites. HSTS preload. No legacy TLS or cleartext fallbacks.
Identity
SSO via SAML 2.0 & OIDC. SCIM 2.0 provisioning. MFA mandatory. FIDO2 / passkeys supported. Break-glass accounts logged and reviewed monthly.
Logging
Signed immutable logs (Grafana Loki + append-only S3). SIEM: Elastic Security. 24/7 SOC. Clock sync via NTS.
Secure SDLC
Mandatory code review · SAST (Semgrep) · DAST (ZAP) · SCA (Dependabot) · SBOM per build (CycloneDX) · signed commits & images (Sigstore/cosign). Quarterly external pen-tests by an EU-based, BSI-recognised provider. Public bug bounty.
§6 GDPR & regulatory alignment
- Role. Data Processor on behalf of the Customer. Standard EU DPA available; supports Customer-supplied DPA on request.
- Lawful bases. Customer determines lawful basis for processing in their own role as Controller. Sub-processor list and international transfer schedule published in the DPA.
- Data subject rights. Self-serve export in JSON and DOCX; admin-initiated full export and delete on request. Deletion timelines (including backups) documented in the DPA.
- Incident response. Playbook-driven; 4h customer notification SLA; 72h regulator notification per GDPR Art. 33.
- EU AI Act. WinGrants AI is classified as a limited-risk AI system under the EU AI Act. Output labelling, transparency obligations, and human-oversight requirements are designed in.
- NIS2. Controls aligned with the NIS2 Directive risk-management framework via the ISO 27001 ISMS.
§7 Contractual commitments
- Data Sovereignty Guarantee (Enterprise) — your data never leaves your perimeter. Written into the DPA, audited annually by your IT, verifiable in our open-source codebase.
- Open-source Guarantee (Enterprise) — every model in the chain is open-source. We will never swap in a closed-model dependency without your written approval.
- Exit Guarantee (Enterprise) — on termination you keep the Docker bundle, your database export, every custom prompt, and the model weights. Stand up an offline instance forever.
- IP Sovereignty Guarantee (AI Enablement) — your prompts, your scorers, your database are yours. Export anytime. The methodology you trained leaves with you.
- 100% Satisfaction Guarantee (all plans) — we iterate until your team is producing submission-ready drafts. If we can't get there under our Fair-Use Policy, the unused portion of the licence is refunded.
What auditors usually ask for next: Statement of Applicability (SoA), pen-test summary, sub-processor list, DPA template, and incident-response playbook excerpt. All available under a short mutual NDA. Email aj@wingrants.ai with subject "Procurement security review" and we'll reply within one business day.