Security at OSQR

OSQR is an AI that reads your documents and conversations in order to work for you. That means it is not "zero-knowledge" and we will never pretend otherwise: our servers decrypt your content to answer your requests, then drop it from working memory. The real security questions are the ones we answer below — who can access what, what leaves our systems, what is retained, and what happens when something goes wrong.

• At rest: user content (documents, messages, summaries) is encrypted with AES-256-GCM using per-account data keys, on top of infrastructure-level disk encryption. A raw database leak yields ciphertext, not your documents.
• In transit: TLS 1.2 or higher (TLS 1.3 on modern clients) everywhere.
• Credentials: passwords are hashed, never stored; OAuth tokens are encrypted at rest.

• You, and the people you explicitly add to your workspace. Data is isolated per account.
• OSQR staff do not browse user content in normal operation. Any access is purpose-bound, restricted, and logged. For child accounts, every access by anyone at OSQR is written to a parent-visible, hash-chained log before a word is read.
• We test on our own account, not yours. When something has to be proven in the live product — a phone line that answers, a form that saves — we do it in an OSQR-owned test workspace that we create, use and clean up ourselves. Your workspace is not a test fixture.
• AI model providers (Anthropic, OpenAI, Google, xAI, Groq) receive only the text a request needs, under agreements that prohibit training on it. Full list, with data types and locations: subprocessors.
• No training, no selling, no ads. OSQR does not train AI models on your data, does not sell it, and does not build advertising profiles from it.

For contractors using OSQR on plans, takeoffs, and bids: your plan sets and estimates are workspace-isolated and are never used to train models or shared across customers. Where our estimating service includes human-in-the-loop processing, that access is part of the service and disclosed — not hidden. If you license specific documents to us in writing (for example, historical takeoffs used to grade accuracy against your own numbers), that written license governs exactly what we may do, and it never includes model training.

• Hosting: Vercel (US) · Database: Neon PostgreSQL (AWS) · Files: AWS S3 — all US-based.
• Payments: Stripe. OSQR does not store card data.
• Data residency is US-only today; EU residency is on the roadmap.

• Your data is retained until you delete it, and deletion is something you perform rather than something that happens to you: the "Burn It" button in Settings destroys your encryption keys first — which makes vault content unreadable immediately — then deletes your account, your workspace and the records inside it, including website and text conversations, voicemail transcripts and call transcripts.
• What closing your account does not do, today. Cancelling a subscription stops the billing and, for a phone line, silences the line at the end of the paid period and hands the number back — it does not erase your content. Your data stays until you press Burn It or ask us to remove it, which we then do by hand. A scheduled purge on termination is on the roadmap and is not built; we would rather say so than print a number we do not enforce.
• Some retention is enforced by automated jobs, not by hand: audit logs pruned at one year, IP addresses redacted at 90 days, model cost logs at 90 days. Conversations, calls, voicemails and contacts are not on a clock — nothing sweeps them.
• What outlives Burn It, on purpose or by plumbing. The do-not-contact list survives deliberately, so that a person who replied STOP can never be resurrected by a fresh account. Contact records are stored without a link back to the workspace, so they are left behind rather than removed with it. And where a business rents a phone number through us, the number itself and any voicemail audio held by the telephony company are released on request rather than by that button — tell us and we will do it, and we will clear the contact rows at the same time.
• Our database provider keeps its own short-term encrypted backups on its own schedule, and deleted rows age out of those with time. We do not keep an internal archive, a cold-storage copy of our own, or a ghost copy for analytics.

• We will notify affected customers within 72 hours of discovering a breach affecting personal data — the same commitment published in our Terms, DPA, and transparency report.
• Found a vulnerability? Tell us: info@osqr.ai (see /.well-known/security.txt). We take coordinated disclosure seriously and will respond within 24 hours.

• No SOC 2 certification yet. A readiness program is underway (compliance platform, then a Type II observation period), with an independent penetration test planned for 2026. Our 15-policy internal control set is available for review on request.
• Not HIPAA compliant — do not upload PHI.
• No native MFA yet — Google sign-in carries Google's MFA; recovery codes protect encrypted-data access; TOTP is on the roadmap.
• We would rather tell you this on page one than have your security team find it on page ten. A small company earns trust with honesty and architecture, not with logos it does not have.