Technology

One verification engine, built to be checked.

How Tumaini reaches an answer: which parts are AI, which are cryptography, which are rules — what has been measured, what is simulated, and what does not exist yet.

Architecture

Three experiences, one engine.

Assist, Trust and Core are front doors to the same API and the same trust-core library. There is no second backend: a person checking a voice note and a bank calling the API reach the same rules.

Experiences

Tumaini Assist

People · web today, mobile companion in development

Tumaini Trust

Organisations · institution portal

Tumaini Core

Developers · public API, sandbox

API layer · Next.js on Vercel

Sessions, tenant isolation, origin checks, rate limits, upload sniffing. Every experience calls the same routes.

The verification engine

trust-core (TypeScript, pure, unit-tested)

Signed receipts · trust registry checks · EN/AR request extraction · deterministic risk policy · protected-action state machine · contextual-signal merge (can only add caution) · evidence signing

Media service (Python, CPU, private)

Voice deepfake model · transcription · C2PA verify & sign

Contextual analysis adapter

Gemini via provider-agnostic interface · consent, redaction, schema validation · not enabled

PostgreSQL with row-level security

Trust registry, receipts, cases, protected actions and a hash-chained audit log. A trigger re-checks every protected-action transition.

Architecture of the running prototype. The dashed box is built and tested but switched off until credentials are supplied and validated.

Capability ledger

What is real, what is simulated, what is not built.

“Implemented” means it runs on the live system and has automated tests. Nothing below is promoted without evidence.

Tumaini capability ledger
Area	Capability	Status	Notes
Media analysis	Voice deepfake detection	Implemented	MMS-300M-AntiDeepfake on CPU, calibrated three-way labels.
Media analysis	Voice transcription	Implemented	faster-whisper base; errors are common, so users can describe the request themselves.
Media analysis	Image manipulation detection	In research	Candidate models under evaluation. Nothing is advertised until validated.
Media analysis	Video deepfake detection	In research	Candidate models under evaluation. Not implemented.
Provenance	C2PA verification (images, video)	Implemented	Official C2PA SDK; demo trust anchors only.
Provenance	C2PA signing for institutions	Implemented	Demo certificates; no trusted timestamp.
Institutional identity	Signed authorisation receipts	Implemented	Ed25519 JWS with content hash, audience, exact scope, single use, revocation.
Institutional identity	Trust registry	Implemented	Demo registry; not an official trust list.
Context	Request extraction (EN/AR)	Implemented	Deterministic rules: money, codes, links, urgency, secrecy, new numbers.
Context	Gemini contextual analysis	Built, not enabled	Provider-agnostic adapter, schemas, limits and consent are built. Disabled until credentials are supplied and tested.
Decision	Deterministic risk policy	Implemented	AI can only add caution; it can never approve.
Decision	Protected-action state machine	Implemented	Enforced in application code and again by a PostgreSQL trigger.
Decision	Trusted-contact confirmation	Simulated	Real one-time-token flow; delivery to the contact's phone is simulated on screen.
Evidence	Signed evidence reports	Implemented	Case reports signed with Tumaini's evidence key; verifiable via API.
Evidence	Hash-chained audit log	Implemented	Per-tenant, tamper-evident.
Organisations	Institution portal	Implemented	Issue, revoke, keys, verification activity, incidents, audit.
Organisations	Webhooks	Planned	Not built.
People	Web verification experience	Implemented	Voice, message, image and video checks in the browser.
People	Tumaini Assist mobile app	Planned	Screen system designed; no native app exists.
People	iOS / Android share integration	Planned	Strategy documented; requires native app and on-device testing.
Payments	Money movement	Simulated	No real funds move. Banks and people are fictional.

Detection, provenance, protection

Three kinds of evidence, never confused.

Detection estimates. Provenance proves. Protection enforces. Tumaini keeps them separate and tells you which is which.

Detection

Probabilistic

Does this look synthetic?

An open pretrained model scores the voice note. It can be wrong in both directions, so it can raise caution but never approve.

GenuineInconclusiveSynthetic

Model: MMS-300M-AntiDeepfake
Runs on: CPU · no paid API
Thresholds: < 0.172 · > 0.859
Calibrated on: Held-out speakers

Scores on the demo files

Synthetic voice note (TTS)0.9985 Likely synthetic
Public-domain human reading0.002 Likely genuine

Easy cases — a “likely genuine” result never identifies the speaker.

Provenance

Cryptographic

Can the source prove it sent this?

Registered organisations sign their messages. Each check is shown — and a missing signature means “unknown”, never “fake”.

Signed authorisation · Falcon Demo Bank

Ed25519 · JWS

passed
SignatureVerified
passed
IssuerRegistered · active
passed
Message textUnchanged
passed
ExpiryValid for 24 h
passed
AuthorisesAED 150.00 → AE07…3456
information
UseOnce only

Change one character of the text or the amount and verification fails. Fictional bank; real Ed25519 signatures.

Protected action

A warning can be ignored. A hold cannot be tapped through.

Every transition is checked in application code and again by a PostgreSQL trigger. RELEASED requires a confirmed check from a contact registered before the message arrived, or a valid signed authorisation for exactly that amount and account.

1CREATED
2ASSESSING
3HELD
4PENDING
5VERIFIED
6RELEASED

AI output can add caution — never approve.
If the AI service is down, holds stay.
A message cannot choose who verifies it.
A signature authorises one action, once.

Detection

Question: Does this look synthetic?
Nature: Probabilistic score
When absent: Inconclusive — verify
Role in Tumaini: Adds caution

Provenance

Question: Can the source prove it sent this?
Nature: Cryptographic check
When absent: Unknown — not evidence of a fake
Role in Tumaini: Can satisfy a check for one exact action

Detection compared with provenance
	Detection	Provenance
Question	Does this look synthetic?	Can the source prove it sent this?
Nature	Probabilistic score	Cryptographic check
When absent	Inconclusive — verify	Unknown — not evidence of a fake
Role in Tumaini	Adds caution	Can satisfy a check for one exact action

Row-level security

Each visitor’s workspace is isolated inside PostgreSQL, not only in code.

Keys at rest

Issuer keys are AES-256-GCM encrypted; revoked keys are destroyed.

Untrusted inputs

Uploads are size-limited and type-sniffed; model output is treated as data.

Explainable

Every outcome lists the checks that passed or failed, in English and Arabic.

Security model and threat analysis

Evaluation and calibration

Tested. Measured. Transparent.

Real metrics, readable denominators and the limitations that matter.

Model. MMS-300M-AntiDeepfake (NII Yamagishi Lab), CPU inference.
Thresholds. Likely genuine below 0.172 · likely synthetic above 0.859 · inconclusive in between — frozen on a speaker-disjoint In-the-Wild calibration split (600/600).

English real-world speech (In-the-Wild, unseen speakers)

2.9%

Equal error rate. n = 1,000 genuine / 1,000 fake. Flagged synthetic: precision 99.9% (913/914), recall 91.3%.

Arabic speech (ArAD, voice conversion)

2.85%

Equal error rate. n = 484 genuine / 1,000 fake. Levantine dialect, short clips; Gulf dialects and Arabic TTS not tested.

WhatsApp-like audio (Opus 16 kbit/s, English)

6.1%

Equal error rate. n = 500 / 500. 24.6% of genuine clips come back “inconclusive”; 5% are wrongly flagged.

Read this before the numbers

These figures describe the model on public datasets of pre-2023 fakes, not real-world accuracy against modern voice cloning. Tumaini’s protection does not depend on the detector: payments are held until independent verification whatever the score.

Why not the best-known model?

AASIST scores 1.72% on its own benchmark but 39.6% on real-world English and 42.1% on Arabic — near chance. We selected on data the models were not trained on, and published the failure.

Model weights: CC BY-NC-SA 4.0 (non-commercial). Suitable for this research prototype; a commercial launch needs a licensed or self-trained model.

Full evaluation report

Contextual intelligence

Gemini reads the situation. It does not judge authenticity.

A multimodal model is good at understanding what a message, picture or recording is asking for. It is not a validated deepfake detector, and Tumaini never treats it as one.

Live status

Checking live status…

The adapter, schemas, limits and consent flow are built and unit-tested. It is switched on only after credentials are supplied and authentication, sample runs, schema validation, latency, cost and fallback have been tested. No Gemini output is shown anywhere on this site until then.

Gemini may

Describe relevant content
Extract claims and claimed identities
Find requests for money, codes or personal data
Extract web addresses
Spot urgency, secrecy and other pressure
Help explain evidence in English and Arabic

Gemini must not

Authorise a payment
Declare a signature valid
Override C2PA verification
Override a trusted-contact check
Release a protected action
Turn an uncertain result into certainty

How that is enforced

Output must match a strict schema; anything else is discarded.
Signals merge through a pure function that can only add caution — tested to never change a decision or remove a check.
Per-request consent; card numbers, IBANs, codes and emails are removed first.
Timeouts, one retry, per-user and daily caps, and a per-request cost ceiling.
Content and responses are not stored; only metadata is audited.
If the provider is down or unset, every other check still works.

Image and video

Not advertised until validated.

Tumaini checks images and video for C2PA Content Credentials today. Manipulation and deepfake detection for images and video is in research: candidate pretrained models are being assessed, and none is integrated.

Before any detector is integrated, it must show

A licence that permits our use
Published, downloadable weights
CPU-feasible inference
Results on independent data, not its training set
Robustness to compression and re-encoding
Measured false-positive and false-negative rates
Generalisation to unseen generators
An acceptable operating cost

Provenance and detection answer different questions. A valid C2PA manifest says who signed a file and whether it changed; it says nothing about whether the scene is real. A missing manifest is “unknown”, never “fake”. The same discipline applied to voice: the best-known benchmark model scored near chance on real-world audio, so a different model was selected on data none of them were trained on.

Privacy and security

Private by default.

Minimal collection
Uploads are analysed and discarded; the app keeps a SHA-256 fingerprint and the result. 4 MB limit.
Tenant isolation
Each visitor's workspace and each organisation are isolated by PostgreSQL row-level security, not only in code.
No covert monitoring
Tumaini only sees what a person chooses to share. The mobile design uses user-initiated sharing, never screen reading.
Keys at rest
Issuer keys are AES-256-GCM encrypted under a server-held key and destroyed on revocation. Production needs HSM/KMS.
External providers
Nothing is sent to an external AI provider without explicit, per-request consent and a clear disclosure.
Never credentials
Tumaini never asks for passwords, PINs or one-time codes, and strips them from text sent to models.

Security model and threat analysis Privacy notice

API integration

The same engine, one call away.

Verify a signed communication at the moment of payment, issue sandbox receipts, verify evidence reports. A working integration with a fictional bank calls the live backend.

Developer documentation Bank integration example

Known limitations

What Tumaini cannot do yet.

Voice-model accuracy is measured on public datasets of pre-2023 fakes; modern cloning may evade it. The weights are licensed for non-commercial use only.
No image or video deepfake detection. No native mobile app or share extension.
Transcription makes mistakes, especially in Arabic; rule-based extraction can miss unusual phrasing.
The trust registry and C2PA trust anchors are demo-only. Organisations and money are fictional.
Trusted-contact delivery is simulated on screen. Gemini contextual analysis is not enabled.
The full limitations list