Incode Frontier AI Lab
AI that evolves faster than fraud
We build our own foundation models for identity, trained on billions of real verifications and stress-tested against the newest gen-AI attacks.
How we adapt to new fraud
Three models, one adaptive defense
Identity Vision-Language Model
Reads faces and documents across 200+ regions to catch tampering, synthetics, and deepfakes.
Fraud Large Language Model
Reads behavior, device, and transaction patterns in real time to expose hidden fraud intent.
Reasoning Agents
Fuse every model and signal into one context-aware risk decision.
What our models learn from
The data engine
Labeling pipelines
200+ human labelers review millions of records to train and score every model.
Synthetic data
120+ generation tools manufacture rare attacks: tampered documents, presentation attacks, deepfakes.
Fraud Lab
A red team that replays real-world attacks against our own models, continuously.
What runs under the hood
The model catalog
Download the full overview (PDF)Detects faces, builds robust embeddings, and matches identities at scale, improving continuously through hard-case mining.
- NIST #1 for facial recognition
- 1:1 and 1:N NIST certified
- FIDO Face certified
- DHS RIVTD benchmarks met
Tells real people and physical documents apart from spoofs and replays, using spatial, temporal, and device-aware signals.
- First passive liveness certified to market
- iBeta ISO 30107-3 PAD Level 2
Detects and blocks AI-generated fraud: deepfakes, face swaps, injected media, and synthetic identities.
- #2 in the ICCV 2025 DeepID Challenge
- #1 in deepfake attack detection, Hochschule Darmstadt
Calibrated age estimation with uncertainty bounds and fairness constraints; edge cases route to secondary verification.
- NIST top 3 for lowest average error
- NIST fastest response time
- ACCS accredited under PAS 1296
Classifies 4,600+ document types, validates OCR, MRZ, and barcodes, and scores authenticity with active learning.
- Evaluated on global document datasets across 200+ regions
Fuses model outputs, network intelligence, and AI risk agents into a single real-time decision.
- 250+ signals fused per identity check
Blocks injected or emulated environments and flags scripted, non-human interaction patterns.
- Detects emulators, virtual cameras, hardware spoofing, and automation
How we govern AI
Responsible AI, by design
Data practices
Purpose-limited, minimized, encrypted, with regional options.
Access & security
Role-based controls, secure SDLC, HSM keys, full audit logs.
Dataset quality
Curated, balanced datasets with continuous QA.
Model development
Reproducible pipelines and versioned training.
Fairness & bias
Bias testing across demographics, with remediation.
Deployment controls
Staged rollouts, canary checks, kill-switches.
Monitoring & feedback
Drift detection and fraud-focused production alerts.
Retention & deletion
Configurable retention with verified deletion.
Incident & continuity
24/7 monitoring and disaster-recovery readiness.
Compliance
SOC 2, ISO, GDPR, CCPA, LGPD, plus a public Trust Center.
What AI technology does Incode use for identity verification?
Incode runs its own frontier AI lab. Facial recognition, liveness, document analysis, and deepfake detection are all custom models, with no third-party AI components.
What is liveness detection?
Technology that confirms a biometric scan captures a real, physically present person, not a photo, replay, or deepfake. Incode's passive liveness works from video frames alone, with no user actions required.
How does Incode compare to off-the-shelf verification APIs?
Incode's models are trained on a proprietary dataset of 4.1B+ identity events and consistently rank at the top of public benchmarks. Off-the-shelf APIs use generalized models that aren't optimized for fraud at scale.
What role does OCR play in document verification?
OCR extracts structured data from identity documents: name, date of birth, document number, expiry. AI post-processing validates formats, cross-references fields, and flags inconsistencies.
How does Incode ensure accuracy across ethnicities and skin tones?
Models are trained on globally diverse datasets and tested against NIST FRTE benchmarks that specifically evaluate demographic fairness, with consistent accuracy across all tested groups.
What's next
Put frontier AI to work
See the models run on your own flows, with your own documents and users.