Is this real, or was it generated?
Verafai scans video, audio, and images for the fingerprints of AI generation — and tells you exactly what it found, in plain language, in under five seconds.
Run a live analysis on a sample file.
Pick a sample below and run it through the same pipeline that powers Verafai — scanning, scoring, and explaining, right in your browser.
Generative AI now produces near-perfect fakes.
Diffusion models, voice cloning, and face-swap tools are good enough to fool a human on casual viewing — and every detector trained on last year's generator quietly goes stale against this year's.
Banks running identity verification, insurers reviewing claims, dating platforms fighting catfishing, newsrooms verifying footage, and courts protecting evidence integrity — all lack a dependable "is this real?" check.
+New generator models ship every month, and each one degrades existing detector accuracy further. This is a moving target, not a solved problem — and most detectors are static models that decay the day they ship.
+Most detectors handle only one modality — just video, or just audio — and return a bare score with no explanation of what triggered it. That's not enough for a compliance team that has to justify a decision.
+In KYC and legal contexts, a missed fake lets fraud through — and a false positive wrongly blocks a real customer. Both failure modes carry direct financial and reputational cost.
+Multi-modal detection that retrains as fast as the fakes evolve.
Verafai combines video, audio, and image classifiers with a continuous retraining pipeline that ingests new synthetic-media samples the moment new generators appear — so accuracy doesn't decay over time.
Ingest
Upload a file through the web app, or stream media directly via the API.
Analyze
Multi-modal models scan for generation artifacts — frequency-domain inconsistencies, lip-sync mismatch, unnatural blinking, voice-clone spectral signatures.
Score
Returns a confidence score from 0–100% likelihood of being synthetic, with flagged timestamps and regions.
Explain
A plain-language explanation of exactly what triggered the flag — e.g. "audio pitch inconsistency at 0:42–0:47."
Retrain
Confirmed fake and real samples continuously feed back into retraining, so the model keeps pace with new generators.
Purpose-built models for each modality, not a generic classifier.
Detection is fundamentally an arms race. Verafai is built from day one to retrain continuously against new generators, rather than shipping a static model that decays.
Video Artifact Classifier
CNN/transformer-based model detecting blending boundaries, unnatural motion, and lighting inconsistency frame-by-frame.
Frame-Level DetectionAudio Spectral Classifier
Detects voice-cloning artifacts in pitch, prosody, and frequency signatures that don't occur in natural speech.
Spectral AnalysisImage Frequency-Domain Model
Detects diffusion and GAN fingerprints invisible to the human eye, using frequency-domain analysis.
GAN FingerprintingExplainability Module
Generates human-readable rationale from model attention and activation maps — not just a raw score.
Attention MapsContinuous Retraining Pipeline
An always-on MLOps loop that retrains and redeploys models as new generator families emerge — the core defensibility of the product.
Core DefensibilityModel Changelog
Every model version ships with a changelog showing exactly which generator families it can detect.
Versioned DeploymentOne of these faces doesn't exist.
Drag the divider. Verafai's frequency-domain model catches the fingerprint your eyes can't — blending seams, texture mismatch, and inconsistent lighting on the synthetic side.
Everything a compliance or trust & safety team needs to act on a result.
Upload & Analyze
Drag-and-drop upload for video, audio, or image files, with a live progress indicator during analysis.
Explainability Layer
Heatmap overlay on flagged frames, timestamped audio flags with waveform annotation, and a human-readable summary.
REST API
Programmatic upload and analysis for KYC, claims, and moderation pipelines, with webhook callbacks on completion.
Case & History Dashboard
Every analysis is stored per user or org, with an exportable PDF report for compliance and legal use.
API Key Management
Issue and rotate API keys with a built-in usage dashboard for every integration.
Live-Stream Detection Phase 2
Real-time analysis for video calls, so a live KYC session can be verified as it happens.
Cloud and GPU infrastructure built for continuous retraining.
Video and audio inference are GPU-heavy, and the retraining loop never stops. This is a genuine, ongoing compute workload — not an LLM wrapper.
How We Use AWS
We use AWS to host our REST API (API Gateway + Lambda), run GPU-backed inference for video, audio, and image models (EC2 G5/P-series or SageMaker endpoints), store uploaded media and training datasets in S3, queue large video jobs with SQS, and deliver results via CloudFront. User accounts, analysis history, and API keys live in RDS or DynamoDB.
How We Use NVIDIA
We plan to use NVIDIA technologies to accelerate model training and inference for our computer-vision and audio classifiers, optimize real-time throughput for video analysis, and scale GPU capacity as retraining frequency increases with new generator families.
Focused on the industries with the most to lose from a fake.
Target Users
Primary (MVP): Fintech and KYC teams verifying identity documents and selfie-video liveness checks.
Secondary: Insurance fraud investigators reviewing claim media.
Tertiary: Dating platforms, trust & safety teams, and independent newsrooms verifying footage.
Revenue Model & Growth
Subscription pricing by monthly analysis volume, with a free trial tier for evaluation and custom enterprise contracts for unlimited-volume, SLA-backed deployments. Beta goal: 20 paying pilot customers within 3 months of launch.
The window to build a defensible detector is closing.
Early, and moving fast.
MVP In Development
Phase 1 build underway: web upload, video and image detection, basic dashboard, and report export.
Pilot Outreach
Currently in conversation with fintech and KYC teams for early pilot access ahead of Phase 1 completion.
Founder-Led Build
Built by a technical founding team with direct experience in computer vision and applied ML.
The founders building Verafai.
A small, technical founding team with direct hands-on experience across computer vision, applied ML, and product engineering.
Mbam Joseph
Leads product, strategy, and pilot partnerships — the primary point of contact for KYC and fintech pilot teams.
David Adeyemi
Leads the detection pipeline and infrastructure — model architecture, retraining loop, and API.
Three phases, twenty weeks.
MVP
Web upload, video and image detection, basic case dashboard, results and report export.
API & Audio
Audio detection, full REST API, and version 1 of the continuous retraining pipeline.
Scale & Compliance
Live-stream detection, batch processing for large media archives, and enterprise compliance controls.
See what Verafai flags in your own media.
Join the pilot waitlist or request a live demo — we're onboarding fintech and KYC teams first.