Research prototype. Synthetic data and simulated payments only — not an official or production security service. Limitations

Pilot

A 90-day pilot with one institution

A proposal. No pilot partnership exists yet; all targets below are targets, not results.

Partner and scope

Primary: a UAE bank or payment provider, on one high-risk flow — transfers to newly added beneficiaries above a threshold — for opted-in retail customers. Alternative: a government entity signing its official SMS and email notices so residents can verify them, with an impersonation-report queue.

Phases

WeeksWork
1–3Governance, security review, data-protection impact assessment, threat model refresh, integration design, success criteria agreed.
4–6Shadow mode: Tumaini scores and logs decisions; nothing is blocked. Measure false-hold rate and latency.
7–11Live for opted-in customers: holds, trusted-contact checks via the bank’s own channels, signed bank messages, report routing.
12–13Independent evaluation; go / extend / stop decision.

Measures (targets)

  • Share of scam-pattern transfers held before release; confirmed fraud losses prevented in the pilot cohort.
  • False-hold rate and median added time for genuine transfers (target: low enough that opted-in customers stay opted in).
  • Verification completion time for trusted-contact checks; share of signed messages verified by recipients.
  • Detector performance on pilot audio with consent, reported separately for Arabic and English.

Approvals and controls

  • Institution security and risk sign-off; regulator engagement as the partner deems appropriate (e.g. CBUAE expectations for banks; TDRA for telecom channels).
  • PDPL (Federal Decree-Law No. 45 of 2021) considerations, data residency in the UAE, minimised data, no voiceprints, retention limits, consent for audio analysis.
  • HSM-backed issuer keys; SSO + MFA for staff; penetration test before live phase.

Rollback

A single feature flag returns the flow to the bank’s existing process; holds fail open to the bank’s standard manual review, never to automatic release. Triggers: false-hold rate above the agreed ceiling, availability below target, or any security incident.

Expansion conditions

Targets met, no unresolved high-severity findings, customer satisfaction maintained, and a governance model for the trust registry agreed with the relevant authorities.