证明是真人。不是深度伪造。不是大型语言模型。

Didit is infrastructure for identity and fraud, the platform we wished existed when we were building products ourselves: open, flexible, and developer-friendly, so it works as a real part of your stack instead of a black box you integrate around.

One API covers verifying people (KYC, know your customer), verifying businesses (KYB, know your business), screening crypto wallets (KYT, know your transaction), and monitoring transactions in real time, on a stack built to be:

• Fast, sub-2-second p99 on every session
• Reliable, in production with 1,500+ companies across 220+ countries
• Secure, SOC 2 Type 1, ISO 27001, GDPR-native, and formally attested by Spain's financial regulator as safer than verifying someone in person

The footprint underneath: 14,000+ document types in 48+ languages, 1,000+ data sources, and 200+ fraud signals on every session. The Didit infrastructure dynamically learns from every session and gets better every day.

A short challenge that proves three things at once:

• A real person is on the other side, captured in real time, not a stored photo, replay, or video file.
• The person is not a printout, screen, silicone mask, or AI-generated face.
• The person is unique, they haven't already passed the same gate on your platform.

The first two are called liveness and Presentation Attack Detection (PAD). The third is 1:N face deduplication. Together they're what most platforms now mean when they say "proof of human".

Two reasons.

• Generative AI shipped. A teenager with a free image model can mint a photorealistic face in seconds. Headless agents can spin up thousands of accounts overnight. Old defences (a CAPTCHA, a phone number, an email) don't bind to a real person anymore.
• Platforms are accountable for what humans on them say. The EU Digital Services Act, the UK Online Safety Act, and several state-level rules treat platforms as responsible for harms caused by inauthentic behaviour. "We didn't know" stopped being a defence in 2024.

The full flow normally takes under 30 seconds end-to-end, pick up the ID, snap the document, snap the selfie, done. That is the fastest in the market. Legacy KYC providers usually take more than 90 seconds for the same flow.

On the back end, Didit returns the result in under two seconds at p99, measured from the moment the user finishes the selfie to the moment your webhook fires. Mobile capture is tuned for slow phones and slow networks: progressive image compression, lazy software development kit load, and a one-tap hand-off from desktop to phone via QR code if the user starts on web.

The model is trained on real captures vs. every known spoof class, print, screen replay, mask, morph, Generative Adversarial Network (GAN) portrait, diffusion render. Real captures carry tiny artefacts no spoof has: micro-motion of the head, sub-pixel skin texture, reflections that match the device's flash spectrum, depth cues from a single passive frame.

The iBeta Level 1 PAD certification, the bar the US National Institute of Standards and Technology cites, measures exactly this performance, and Didit holds it across every passive-liveness session.

Every session lands on one of seven clear statuses, so your code always knows what to do:

• Approved, every check passed. Move the user forward.
• Declined, one or more checks failed. You can allow the user to resubmit the specific failed step (for example, re-take the selfie) without re-running the whole flow.
• In Review, flagged for compliance review. Open the case in the console, see every signal, decide approve or decline.
• In Progress, user is mid-flow.
• Not Started, link sent, user has not opened it yet. Send a reminder if it sits too long.
• Abandoned, user opened the link but did not finish in time. Re-engage or expire.
• Expired, the session link aged out. Create a new session.

A signed webhook fires on every status change, so your database always stays in sync. Abandoned and declined sessions are free.

Production data is processed and stored in the European Union by default, on Amazon Web Services. Enterprise contracts can request alternative regions for jurisdictions whose regulators require it.

Encryption everywhere. AES-256 at rest across every database, object store, and backup. Transport Layer Security 1.3 in transit on every API call, webhook, and Business Console session. Biometric data is encrypted under a separate Customer Master Key.

Retention is yours to control. Default retention is indefinite (unlimited) unless you configure shorter, between 30 days and 10 years per application, and you can delete any individual session at any time from the dashboard or the API.

Certifications: SOC 2 Type 1 (Type 2 audit in progress), ISO/IEC 27001:2022, iBeta Level 1 PAD, and a public attestation from Spain''s Tesoro / SEPBLAC / CNMV that Didit''s remote identity verification is safer than verifying someone in person. Full report at /security-compliance.

Didit ships compliant by default for the regulators that matter to identity infrastructure:

• GDPR + UK GDPR, controller / processor split, full Data Processing Agreement published, lead supervisory authority named (Spain''s AEPD).
• AMLD6 + EU AML Single Rulebook, 1,300+ sanctions, politically exposed person, and adverse-media lists screened in real time.
• eIDAS 2.0, EU Digital Identity Wallet aligned; reusable-identity ready.
• MiCA (Markets in Crypto-Assets), ready for crypto on-ramps, exchanges, and custodians.
• DORA, Digital Operational Resilience Act, EU financial-services operational resilience.
• BIPA, CUBI, Washington HB 1493, CCPA / CPRA, US biometric privacy (Illinois, Texas, Washington) and California consumer privacy.
• UK Online Safety Act, age-gating and child-safety obligations.
• FATF Travel Rule, originator and beneficiary data on crypto transfers, IVMS-101 interoperable.

Detailed memo, every certificate, every regulator letter: /security-compliance.

• 60 seconds to a sandbox account at business.didit.me, no credit card.
• 5 minutes to a working verification through Claude Code, Cursor, or any coding agent via our Model Context Protocol (MCP) server.
• A weekend to a production-ready integration with signed-webhook verification, retries, and a remediation flow when a user is declined.

Three integration paths, pick whichever fits your stack:

• Embed natively with our Web, iOS, Android, React Native, or Flutter SDK.
• Redirect the user to the hosted verification page, zero SDK.
• Send a link by email, SMS, WhatsApp, or any channel, zero front-end work.

Same dashboard, same billing, same pay-per-success price for all three. Step-by-step guide at docs.didit.me/integration/integration-prompt.

Two paths.

• Hosted workflow, POST /v3/session/ with the workflow_id you built in the Business Console. We host the capture UI, run Liveness + Face Search 1:N, return a signed verdict to your webhook. Five-minute integration.
• Server-to-server, call POST /v3/passive-liveness/ with the selfie you already captured locally, then POST /v3/face-search/ for Sybil dedupe.

Full cURL example, response shape, and the agent-pastable prompt are above. The Model Context Protocol (MCP) server speaks both surfaces to Claude, Cursor, and any other agent.

Data minimisation by default. The selfie is processed in memory; the verdict, liveness score, and an irreversible face template are persisted; the raw image is deleted unless retention is explicitly enabled. The face template is a one-way hash, you cannot reconstruct the underlying face from it.

We never sell, share, or train third-party models on a customer's biometric data. The full data-processing terms live at didit.me/terms/business; the privacy notice the end user sees is at didit.me/terms/verification-privacy-notice.

The session returns one of four outcomes:

• Approved, clean human. Issue the Reusable Credential and grant access.
• Declined, clear spoof signal (print, screen, mask, generated face). Block.
• In Review, high-similarity match in Face Search 1:N or a low liveness score. Hold the surface, route to your ops queue. The Business Console gives the analyst the selfie, the match, and the score.
• Not Finished, user abandoned the capture. Safe to re-prompt.

A signed webhook lands at every state change, so you only render against the latest verdict.