Built for the regulator, proven in the region
A fraud-intervention agent speaks to your customers in their most anxious moment. That only works inside a security posture banks can sign off: UAE data residency, PDPL-aligned processing, deterministic and auditable decisions, and a hard rule that the agent never asks for credentials.
Regulatory posture
Processing designed around Federal Decree-Law No. 45 of 2021: minimal personal data, documented lawful basis, in-region storage, and DSR workflows that honour 30-day windows.
Interventions map to Consumer Protection Regulation expectations: immediate protective action on confirmed fraud, full call recording, and same-day regulator-ready case files.
The agent never touches PANs. Card identifiers arrive as tokens and last-4 only, keeping the voice platform entirely outside the PCI cardholder-data environment — the applicable self-assessment is SAQ A, not SAQ D.
This deployment runs in European infrastructure (Frankfurt, eu-central-1), TLS 1.3 in transit and AES-256 at rest. Speech synthesis and transcription are performed by processors in the US and UK — no call audio reaches a model for decision-making. A bank requiring in-country (UAE) processing deploys the same containers inside its own VPC, so transcripts and case data never leave its perimeter.
Sessions are self-hosted on our own infrastructure (no third-party identity processor). Verified email sign-in, brute-force lockout, a 15-minute idle timeout and a hard 8-hour session ceiling, so no credential token outlives a working day. Roles are provisioned by invitation — there is no public sign-up.
Every turn, delivery and outcome is a sha256-chained record. Any edit breaks the chain and the built-in verifier names the exact broken row. Organization id is sealed into each link.
Bring-your-own ElevenLabs keys are AES-256-GCM encrypted at rest, decrypted only in-process for an upstream call, and never displayed beyond a masked form.
Platform-key voice usage is metered per workspace daily; interventions consume prepaid credits (402 at zero); the fire endpoint is rate-limited and idempotent.
Platform controls
TLS 1.3 in transit, AES-256 at rest, and HMAC-SHA256 signing on every webhook so consumers verify origin and integrity of each event.
Voice model credentials live only in the server runtime. The browser receives rendered audio — never keys, never raw model access.
Runs inside the bank's own tenancy behind private networking. Telephony and model endpoints are allow-listed; egress is logged and deny-by-default.
Every turn stores the signed policy hash, inputs and outcome — any intervention decision can be replayed byte-for-byte for an auditor or the central bank.
Voice personas are created from written, revocable consent only, watermarked at synthesis, and never reused across institutions.
Transcripts are pseudonymised after case closure; raw audio retention is configurable down to zero once the audit hash is written.
The one rule that defines the product
The agent never asks for PINs, passwords, OTPs or full card numbers. This is not a prompt instruction — it is structural: the classifier has no branch that can produce such a request, and the signed guardrail policy makes any deviation fail closed.
Responsible disclosure
Found something? Report to otemaach@gmail.com (PGP key on request). We acknowledge within 24 hours, triage within 72, and credit researchers who follow scope rules. No legal action for good-faith research that avoids service degradation and never touches customer data.
Encryption: PGP · key ID 0x5ECURE
Preferred languages: EN, AR
Response SLA: 24h ack · 72h triage