Security, data residency, and the build order — in the open.
A regulated-AI buyer should be able to read how we handle data, where it lives, and what they can roll back. This page is the single source of truth. It links to the redactor, the build order, the rollback plan, and our responsible-disclosure policy.
Hosted in Australia. Personal details stripped before any AI call.
By default, every system we build runs on Australian infrastructure (Sydney region). When AI processing is involved, personal details are removed at the call boundary — the offshore model never sees them. This is documented in the build order and verified by our Privacy page.
Vercel
Edge + SSR runs in AWS Sydney region (ap-southeast-2) by default. No global edge caching of personalised content.
Supabase
Postgres, auth, and storage on AWS Sydney region. RLS on every table; append-only triggers on ada_log.
Cloudflare DNS + Email Routing
Sydney region; SPF, DKIM, DMARC enforced on the sending domain.
PII redactor at the boundary
Every text that leaves for an offshore model (DeepSeek chat; OpenAI embeddings) runs through redactPII first. See the source.
Pattern 3 — honest framing
Pattern 3 is a v1 mitigation, not the destination. The public wording reflects what is true, not what is convenient.
Australian Privacy Act (Dec 2026)
Every system includes the transparency notices, human-review points, and audit logging required for the new Automated Decision-Making rules.
What we strip before any text leaves for an offshore model.
app/src/lib/ai/redact.ts is the redactor. Conservative by design — false positives are cheaper than a PII leak. 11 PII kinds covered today; 30-sample test gates any push.
Safe-fail. If a single turn produces more than 5 redactions, Ada refuses to answer and points the visitor to the AI Audit form instead. Every refusal writes a row to the append-only ada_log table.
Australia's AI Ethics Framework — our self-assessment.
Australia's AI Ethics Framework (DISR, 2019) lists 8 principles. Here is how we rate ourselves against each, in plain English. Where we are weak, we say so.
Human, social and environmental well-being
Met
Human-centred values
Met
Fairness
Met
Privacy protection and security
Met (PII redactor; AU residency)
Reliability and safety
Met (75-check regression on ForgeWright)
Transparency and explainability
Met (audit tree + append-only log)
Contestability
Met (human gate; version-bound signoff)
Accountability
Met (named human, versioned signoff, log)
We are not ISO 42001-certified (the international AI management system standard). The ForgeWright process is consistent with it; certification is on the roadmap.
The build order and the rollback plan — in the open.
We publish our engineering and engagement process so a buyer can audit us before signing. The CTO Build Order is the canonical method; the rollback plan is the per-commit revert map for the live site.
AI handles the first 80%. The rest gets a named human.
For real money (> A$10k), legal risk, account-level decisions, anything regulated, NDAs, or contracts — Ada is honest about the limits and hands off to a human. Three handoff paths:
- From the chat widget: click Need a human right now? at the bottom of Ada's chat panel — opens a pre-filled email to finn@goldmanglobal.com.au.
- From Ada in-conversation: Ada will call an escalation tool when the question fits. You get a structured handoff row + an email with the full context.
- From the free AI Audit form: regulated industries go straight to the operator dashboard, bucketed as ACT NOW.
Operator dashboard: /admin/needs-me — 4 priority buckets, top-down in 30 seconds. Auth-gated by Vercel Authentication.
Found a security issue? Tell us first.
We acknowledge reports within 5 business days and aim to triage within 10. We do not currently run a paid bug-bounty programme; we credit researchers in our changelog on a best-effort basis.
Email: security@goldmanglobal.com.au
security.txt: /.well-known/security.txt (RFC 9116).
Scope: goldmanglobal.com.au, all subdomains, and claimdesk247.com.au. Please do not test against any other customer deployment.
Want to talk security before you sign?
The free AI Audit includes a 30-minute security & governance review for regulated industries — we'll walk you through the build order, the rollback plan, and the redactor.