Sovereign by default.
DROZlegal was built with bar councils and ethics committees in the room. Every Canadian firm’s files live in AWS ca-central-1 (Montréal) — stored client data never leaves Canada. AI requests are processed under Anthropic’s commercial terms: nothing trains a model, and inputs and outputs are deleted automatically within approximately 30 days. Privilege is preserved end to end — in writing, in the engagement, in the architecture.
Where we stand, and where we’re headed.
The regulatory frameworks DROZlegal aligns to today, and the independent certifications underway.
SOC 2 Type II
Certification program in progress, covering Security, Availability, Confidentiality and Privacy trust-service criteria. Not yet attested; no report exists yet.
PIPEDA
Federal privacy compliance documentation formalised. Data Processing Addendum available on request, written into your engagement.
Quebec Law 25
Privacy obligations and consent flows for Quebec residents. Bilingual UI strings, French legal terminology, Quebec court rules.
LSO By-Law 9
Trust accounting books match the Law Society of Ontario’s required records, including three-way reconciliation within 25 days of month-end.
LSO Rule 3.4
Conflict-checking inside the CRM uses normalised name, phone-digits-only and exact-email matching for compliant duty-of-loyalty checks.
LSO Rule 3.6
Invoicing logic respects Ontario HST 13% on fees and exempt status on disbursements. Aging reports built in.
ISO 27001
Information-security management certification in progress. Not yet certified; no ISO 27001 audit has been completed.
FIPPA-scoped audit
Immutable per-mutation audit logging tuned for Freedom of Information & Protection of Privacy Act-scoped public-sector workspaces.
The infrastructure that keeps your file yours.
Every technical choice we made, named.
- AWS ca-central-1 (Montréal) deployment — stored data never leaves Canada; AI processing runs under no-training, ~30-day auto-deletion commercial terms.
- AES-256 encryption at rest with keys you can rotate.
- TLS 1.2+ in transit.
- Microsoft Entra ID SSO with MFA enforced by policy.
- Per-org multi-tenancy with org_id scoping on every query.
- RBAC: admin / user / viewer roles, configurable per workspace.
- License kill-switch middleware for instant workspace pause without state loss.
- PII redaction with regex + AI NER before any model call.
- Virus scanning on every upload via ClamAV / Lambda.
- S3 storage with post-write verification.
- Immutable audit logging, per mutation, queryable.
- Metrics service: latency, AI ops, token consumption per request.
- Sentry error tracking and structlog structured logging.
- Redis rate-limiting on outbound model calls.
- EN/FR UI strings end-to-end.
- No files used to train any model — ours, vendors’, anyone’s.
- E&O insurance secured for AI actions before they were enabled.
Why we use retrieval, not megaprompts.
Large context windows degrade model quality through attention dilution. We keep the model focused on the most relevant passages per query — better answers, lower token spend, and zero "lost in the middle" failures.
- pgvector for semantic embeddings, hosted in ca-central-1.
- Voyage AI 1024-dim embeddings.
- Hybrid search — semantic + keyword — via Reciprocal Rank Fusion.
- Per-matter, per-firm and per-user retrieval scoping.
- Chunking strategies tuned per document type (pleadings, contracts, emails, transcripts).
- Re-ranking layer for retrieval quality.
- Citation source preservation through the entire retrieval pipeline.
- Voice and style learning stored as retrievable embeddings, not stuffed into prompts.
Send our security pack to your IT lead. Read it together.
Full security architecture document, sub-processor list and a draft engagement schedule available on request. SOC 2 Type II and ISO 27001 certification are in progress, not yet held.