01
INPUT01
Text + Vision + Sources
Bring the content.
Text, images, claims, and source material enter one unified analysis pipeline.
Multimodal AI · Evidence Intelligence
Don't just detect.
Understand.
VeriLens brings text, images, models, sources, and evidence together to reveal what a piece of content actually means — and why.
Text
/ Vision
/ Evidence
/ Explainability
Content INPUT
Text + image + source material entering the analysis system.
↗
Source
Reference signal
VeriLens engine
NLP
Language
Claim signals
VIS
Vision
Visual signals
Evidence
Connections forming
How VeriLens thinks
From raw content to understanding.
VeriLens does not stop at a prediction. It follows the signal through models, sources, evidence, and relationships until the result becomes understandable.
02
ANALYZE02
Multimodal Intelligence
Read every signal.
Language, visual patterns, metadata, and semantic relationships are analyzed together.
03
VERIFY03
Evidence Retrieval
Find the evidence.
Claims are connected to supporting, contradicting, and contextual evidence.
04
UNDERSTAND04
Explainable Verdict
See why it matters.
Models, confidence, sources, explanations, and relationships become one interpretable result.
Every stage leaves a trace — so the final result is not just a label, but an explanation you can inspect.
Traceable by design
Intelligence layers
One piece of content.
Multiple ways to understand it.
VeriLens combines language, vision, multimodal signals, and machine learning into one analysis system — so no single model has to tell the whole story.
VeriLens intelligence engine
Signals are analyzed independently before they converge.
Analysis active
NLP
LANGUAGE
Read the claim.
Extract claims, entities, numbers, temporal signals, and linguistic patterns from the content.
Claim extraction · TF-IDF · Transformers
VISION
VISION
Inspect the image.
Visual features reveal patterns that text alone cannot see — from image structure to authenticity signals.
CNN · HOG · Vision models
Fusion engine
FUSION
MULTIMODAL
Connect the signals.
Text and visual representations are compared to surface semantic consistency and contradictions.
Embeddings · Similarity · Cross-modal signals
ML
PREDICTION
Measure uncertainty.
Multiple model signals become interpretable predictions with confidence rather than unexplained labels.
LR · Random Forest · XGBoost · Explainability
Signal
Linguistic evidence
Signal
Visual evidence
Output
Interpretable confidence
Different models see different things. VeriLens brings those perspectives together instead of hiding them behind one score.
Multiple signals → one understandingEvidence intelligence
Don't hide the reasoning.
Every important claim can be traced through sources, relationships, relevance, and confidence — turning a prediction into something you can actually investigate.
E
Claim under investigation
“The unemployment rate fell significantly in 2024.”
Evidence mapped
Claim
Claim signal #001
“The unemployment rate fell significantly in 2024.”
Temporal signal 2024
Claim type Numeric
Evidence links 3
Evidence relationships →
Evidence network
Sources connected to the claim
SUPPORTS
National Statistics Office
Annual Labor Market Review
Strength
0.83
CONTRADICTS
Independent Research Archive
Employment Trends Report
Strength
0.71
CONTEXT
Economic Outlook
Quarterly Market Analysis
Strength
0.64
Supporting
Contradicting
Context
Evidence strength combines relevance, source reliability, and relationship confidence.
Claim → Source → EvidenceTraceable
Follow a prediction back to the evidence behind it.
Relational
Supporting and contradicting signals coexist in the graph.
Grounded
Explanations stay connected to retrieved evidence.
Evidence graph
Truth is rarely one-dimensional.
VeriLens maps claims, sources, evidence, and contradictions into a relationship graph — giving you a view of the information surrounding a prediction.
Knowledge graph
Claim relationship map
Supporting
Contradicting
CLAIM
Unemployment fell in 2024
SUPPORTS
Labor Market Review
SOURCE
National Statistics Office
CONTRADICTS
Employment Trends Report
CONTEXT
Economic Outlook
Claims
01
Evidence
04
Relationships
08
The graph keeps the relationships visible instead of flattening everything into a single confidence number.
Explore the evidence network →Explainable intelligence
Every prediction has a reason.
VeriLens exposes the signals behind a model's decision — from important features to supporting evidence — so confidence never has to mean blind trust.
Model explanation
Why this content received its prediction
High risk 78.4%
Model confidence
78.4%
AI-generated probability
Prediction REVIEW
Requires further verification
Evidence and model signals indicate elevated uncertainty.
Feature contribution
Signals influencing the prediction
Claim language0.82
Source consistency0.74
Visual signal0.61
Temporal mismatch0.43
Cross-modal alignment0.31
AI
Explanation
The prediction is primarily influenced by claim language and source consistency, while visual and temporal signals increase the overall risk.
SHAP Feature importance Confidence
Transparent
See which signals influenced the model.
Inspectable
Move from prediction back to evidence.
Human-readable
Turn technical model signals into understandable reasoning.
Model lab
Don't just use a model. understand it.
Compare experiments, inspect evaluation metrics, and understand why a model was selected before it ever reaches a prediction.
Experiment workspace
Model comparison
LIAR2 Evaluation Experiment complete
Candidates
Trained models
Logistic Regression
Text model
Acc
33.5%
F1
26.9%
AUC
0.62
Random Forest
Text model
Acc
31.9%
F1
28.1%
AUC
0.61
DistilBERT
Text model
Acc
29.8%
F1
29.2%
AUC
0.60
HOG + Logistic Regression
Vision model
Acc
72.9%
F1
72.9%
AUC
0.80
Selected experiment
HOG + Logistic Regression
Lightweight vision baseline evaluated on the CIFAKE dataset.
Accuracy
72.9%
Macro F1
72.9%
ROC-AUC
0.80
Evaluation
Confusion matrix
Pred. Real
Pred. Fake
Actual Real
724
276
271
729
Model selection
The selected model provides a lightweight vision baseline with measurable performance and interpretable features. Model choice should be driven by the evaluation context, not accuracy alone.
Train
Feature engineering, preprocessing and model training.
Evaluate
Accuracy, precision, recall, F1, ROC-AUC and confusion matrices.
Compare
Compare experiments before selecting a production candidate.
MLOps & monitoring
A model is never finished.
VeriLens treats models as living systems — tracking predictions, latency, distributions, versions, and signals that can reveal when behavior starts to change.
Production monitor
VeriLens model health
Monitoring active Last 7 days
Predictions
12,482
+8.4% this week
Avg. latency
184 ms
within target
High-risk rate
18.7%
stable
Model version
v0.4.2
production
Prediction distribution
Risk scores across recent predictions
Very low18%
Low27%
Medium31%
High17%
Critical7%
Confidence trend
StableSystem health
Signals worth watching
Prediction latency
184 ms average
Input distribution
No significant drift
High-risk predictions
18.7% of recent traffic
Model version
Productionvl-multimodal-v0.4.2
Dataset v0.4 · Evaluation tracked · Logged
Versioned
Keep datasets, models and experiments traceable.
Observable
Monitor latency, confidence and prediction behavior.
Auditable
Preserve the trail from model version to prediction.
Start with the evidence
Don't just ask if it's true.
Understand the content, inspect the signals, trace the evidence, and see why the model reached its conclusion.
Multimodal AI·Evidence Graph·Explainable ML·Model Monitoring