Check what you see. Know what to do.
A voice note sounds exactly like someone you love and asks for money now. Tumaini checks the message, explains what it is asking in plain Arabic or English, and holds the payment until someone you trust confirms it.
See it work in two minutes.
Every step in this film is a real interaction with the deployed prototype, not a mock-up. A voice note is checked, a signed bank request is verified, a changed amount is blocked, and image credentials are validated.
The film has on-screen captions and no narration, so you can watch it with the sound off.
Fictional people, organisations and payments · Captioned · 2 min 15 s
The decision happens in seconds.
Scams are built for seconds
A voice note that sounds like a family member, an urgent tone, and a request not to call. The person receiving it has seconds to decide and nothing to check it with.
Detectors alone fall behind
Deepfake detectors trail new tools, are trained mostly on English, and return scores people cannot act on. Tumaini does not depend on winning that race.
Real messages cannot prove themselves
A genuine bank or government request has no reliable way to show it is genuine, so real requests get doubted too. Tumaini gives institutions a way to prove it.
Hold the money, not the verdict.
You cannot make a detector perfect. You can make sure a fake on its own is not enough to move money. Tumaini puts the check at the moment of payment and waits for something independent to confirm it.
Check the message
An open anti-spoofing model and a transcript analyse each voice note. The verdict is written in plain words, and an “inconclusive” band appears when the model is not sure.
Understand the ask
Tumaini pulls out what the message wants: a payment, a code, a link, a deadline, and the pressure around it, such as urgency or secrecy. It works in English and Arabic.
Hold the payment
A payment requested through the flow enters a protected state. The AI cannot release it, and the hold stays if the AI is unavailable.
Confirm or cancel
A contact you registered before the message arrived confirms or cancels. Or a signed receipt from the institution matches that exact payment, and it works once. Either way, Tumaini can produce a signed evidence report.
What Tumaini does today.
Voice-note checks you can read
An open anti-spoofing model scores each voice note on a CPU server, with no paid AI service in the loop. Uncertain results show as inconclusive instead of false confidence.
Plain-language verdicts in Arabic and English
The reason and the safe next step are written for the person reading them, so the message is understood at the moment it matters.
Request understanding
Finds payment amounts, codes, links and deadlines in a message, along with the pressure tactics around them, so you see what is actually being asked.
Protected payments
Requested payments are held by a state machine enforced in both the API and the database. Release is refused until a registered contact confirms, or a matching signed receipt is used.
Signed receipts for institutions
Banks and public bodies sign each genuine request with Ed25519. The receipt names the issuer, action, amount, beneficiary and expiry. Change any of them and it fails.
Content Credentials for official media
Checks C2PA Content Credentials on official images and video, and supports signing them for institutions. Missing credentials show as unknown, not as fake.
Evidence you can keep
Signed evidence reports and a hash-chained audit log record what happened and in what order, so a case can be reviewed later.
Built for institutions and developers
A portal to issue receipts, review verifications and respond to incidents, plus a public verification API with a sandbox and a working example integration.
The pitch, on video and in the deck.
The challenge, the evidence, the approach and the limits, presented for the UAE Hackathon 2026, Track 2 Pioneers, Challenge 40: Building Trust in Digital Content.
The video is the pitch as it was presented. Every number matches the evidence section below, with the caveats included.
Pitch video · UAE Hackathon 2026
The same story in ten slides, to read at your own pace or share with your team.
Pitch deck · Canva
Three ways in.
Tumaini Assist · for people
Web checks today, covering voice notes, requests and payment holds. A mobile companion is in development and appears only as concept screens.
Tumaini Trust · for institutions
A portal organised around Prove, Protect and Respond: issue signed receipts, check what is genuine, and respond to impersonation reports.
Tumaini Core · for developers
A public verification API, a sandbox, and an example Node.js integration with a fictional bank. Production API keys are still on the roadmap.
Measured on real-world audio, with the limits shown.
| Test set | EER | What it means |
|---|---|---|
| English, clean audio | 2.90% | 1 in 1,000 genuine clips flagged. 1.3% of fakes labelled likely genuine. |
| English, WhatsApp-like audio | 6.10% | 24.6% of genuine clips come back inconclusive. This is why the band exists. |
| Arabic, clean audio (Levantine) | 2.85% | 5.5% of fakes labelled likely genuine. Gulf dialects were not tested. |
| AASIST, a widely cited open baseline, English real-world audio | 39.6% | Close to chance on the same English real-world audio, which is why we did not use it. |
Each figure is one measurement on one test set under one condition. The test sets range from about 500 to 1,000 clips per class. We do not report a single accuracy number, and a score is not a probability.
The full method, thresholds and caveats are documented in the project’s evaluation report. The detector is licensed for non-commercial research, and a commercial launch would need a licensed or self-trained model.
What we have not built yet.
Image and video deepfakes are not detected yet
Images and video receive Content Credential checks only. Tumaini does not yet judge whether a photo or clip has been manipulated.
Modern voice clones and Gulf dialects were not tested
Our Arabic results come from one public dataset of Levantine voice conversions. We make no claim about general Arabic accuracy.
Payments and delivery are simulated
The prototype does not move real funds or send real messages. Trusted-contact confirmations appear on a simulated phone screen.
Not an official or endorsed service
Tumaini is a research prototype built for the UAE Hackathon 2026. It is not affiliated with or endorsed by any government entity or bank, and no pilot partner exists yet.
Questions people ask us.
Is Tumaini live with a bank?
Not yet. Tumaini is a research prototype with no pilot partner, and we will not claim one. We would welcome a conversation with banks, payment providers and public bodies about what a pilot could look like.
Can I try it now?
Yes. The live demo at tumaini-tau.vercel.app/demo runs on fictional organisations, people and payments. Nothing moves real money.
Does Tumaini tell me whether a message is fake?
It shows what it can see: the audio score with an inconclusive band, the request inside the message, and whether a signed receipt or Content Credential matches. A valid signature shows where a message came from and that it has not changed. It does not prove the message is true. When the evidence is weak, Tumaini says so and keeps the payment on hold.
Who can sign a receipt?
Only institutions registered in the trust registry, using keys that the registry recognises. In the prototype, the registry contains demo data.
See it work, or talk about a pilot.
Start with the live demo. It runs in your browser and uses fictional people, organisations and payments.
If you run a bank, payment provider, public service or community programme and want to explore how protected payments would work for your customers, tell us what you are trying to protect and we will get back to you.
