What it does, how it was built, where it works well, and where it doesn't. We publish methodology and limitations because that's what trust looks like.
WHAT MAKES TUTELIQ DIFFERENT
- EU-resident inference End-to-end. Audio, text, and images stay inside the EU residency boundary.
EU-resident inference
End-to-end. Audio, text, and images stay inside the EU residency boundary.
- No CSAM in training Detection is built compositionally, never from learning the material itself.
No CSAM in training
Detection is built compositionally, never from learning the material itself.
- Domain-expert review Cybercrime detective, forensic psychologists, and criminologists on the panel.
Domain-expert review
Cybercrime detective, forensic psychologists, and criminologists on the panel.
- 50M+ annotated examples Expert-annotated across 32 languages, with cultural calibration.
50M+ annotated examples
Expert-annotated across 32 languages, with cultural calibration.
- Substantiated metrics only We publish numbers when we have a documented benchmark. Otherwise: methodology.
Substantiated metrics only
We publish numbers when we have a documented benchmark. Otherwise: methodology.
Safety
BASE
tuteliq-7b-safety-v1 - custom 7B safety model fine-tuned on 50M+ expert-annotated examples across 32 languages. EU-resident inference on Google Vertex AI.
EVALUATION
Reviewed by a domain panel including a UK cybercrime detective, forensic psychologists, and criminologists. Scenarios cover flash grooming, gaming-platform exploitation, cross-platform migration, and emotional manipulation of vulnerable youth.
STRENGTH
Per-message risk progression. Track escalation across a conversation, not just final-message verdict.
KNOWN LIMITATIONS
Designed cautiously conservative. Favours false positives on ambiguous family or peer interactions over missing real grooming. Pair with the vulnerability-exploitation detector for under-13 audiences.
tuteliq-7b-safety-v1 with crisis-language prompts reviewed by mental-health professionals. EU-resident inference.
Multilingual scenarios reviewed by domain experts, including counter-narrative and reportage cases. Returns evidence quotes for every detection.
Distinguishes ideation from reportage, lyrics, and fiction. The evidence_quotes field surfaces the exact triggering language to human reviewers.
Genuine ideation in subtle or coded language can read similarly to creative writing. Surface evidence to reviewers and apply organisational escalation policy.
tuteliq-7b-safety-v1 with multi-tactic classification covering targeting, exclusion, humiliation, intimidation, and identity-based attacks.
Reviewed against UK Online Safety Act and EU DSA Art. 14 categories.
Detects identity-based harassment (LGBTQ+, neurodivergent, racial) as a distinct vector. Many competitors miss this.
Reciprocal banter inside friend groups can false-positive. Tune sensitivity per surface.
Tuteliq's proprietary forensic ensemble. 25+ analyzers including CFA pattern, GAN frequency, JPEG quantization, lighting consistency, body segmentation, pose estimation, hand anomaly, and age estimation.
ETHICAL FOUNDATION
No actual CSAM is used in training or evaluation. Detection is built compositionally from age estimation, anatomical detection, skin exposure analysis, and benign-context recognition. Not from learning the material.
Catches previously unseen content that hash databases (NCMEC / IWF) cannot match, including newly generated AI imagery. We recommend running both layers: hash matching catches known content, Tuteliq catches unknown.
BENCHMARK
AUC-ROC 1.000, accuracy 99.9%, TPR 99.94% at 1% FPR on a 1,997-image evaluation spanning 19 generative models (DALL-E 3, Midjourney, Flux, Stable Diffusion, Ideogram, Recraft, Seedream, GPT-Image, Google Nano, Wan, and others). Generator attribution top-1 accuracy: 97.34%.
Heavily stylised art and edge swimwear or lingerie cases are routed to "uncertain" rather than "exposed" by design, to preserve human-review judgement.
Fraud
tuteliq-7b-safety-v1 with scam-pattern coverage across romance scams, impersonation, urgency manipulation, and authority spoofing.
Reviewed against bank-AML and dating-platform abuse patterns.
Tracks scam arcs across conversation turns. Catches grooming-style scams that single-message classifiers miss.
Adversarial rewording can reduce recall. Model refresh cadence is regular as new attack patterns emerge.
Conversation-arc analysis on tuteliq-7b-safety-v1. EU-resident inference.
Performance improves substantially with ≥6 message turns of context. Purpose-built for sustained interactions, not one-shot classification.
Single-message evaluation (e.g., a single profile bio) is not the intended use case.
Specialist classifier trained on public mule-recruitment patterns and AML indicators.
Coverage of coded-language rotations through frequent re-training. Financial-crime patterns drift fast.
Newest coded language can lag detection until the next refresh.
Verification
ICAO 9303 MRZ parser with check-digit validation, @vladmandic/face-api (128-dim face descriptors) for selfie-to-document matching, MediaPipe FaceLandmarker liveness with anti-spoofing (challenge-response head turns and blinks).
COVERAGE
Multi-strategy passport OCR, deterministic document field extraction (MRZ check digits, PDF417/AAMVA), Swedish national ID document support, multilingual document handling. No document image is sent to any language model.
SECURITY
Liveness challenges signed with HMAC-SHA256. Sensitive PII (DOB, name, document number) encrypted at rest with AES-256-GCM. No raw document images persisted.
Heavy occlusion (mask + sunglasses) can fail liveness. The system surfaces specific failure reasons rather than a single "rejected" verdict so customers can guide retries.
Analysis
OpenAI Whisper, self-hosted on Google Vertex AI in the EU (no audio leaves the EU residency boundary). Downstream safety classification on tuteliq-7b-safety-v1.
32 languages: English stable, 31 in beta.
Full audio pipeline runs in the EU. Distinct from competitors who route audio through US-based transcription APIs.
Below-room-tone audio (≤ −30 dBFS) and overlapping speakers degrade transcription. Safety classification then runs on the (degraded) transcript.
Want to evaluate on your data?
We share extended benchmarks, confusion matrices and bias audits under NDA with procurement and academic researchers.