Claim 1 [preamble]
A computer-implemented method for content moderation comprising:Claim 1.1
receiving, by one or more processors, a content item submitted to an online platform;Claim 1.2
extracting, by the one or more processors, a plurality of features from the content item using a trained machine learning model;Claim 1.3
generating, by the one or more processors, a classification score based on the extracted features;Claim 1.4
comparing the classification score to a predetermined threshold value;Claim 1.5
in response to the classification score exceeding the threshold, automatically flagging the content item for review.Claim 2
The method of claim 1, wherein the trained machine learning model comprises a convolutional neural network.Claim 3
The method of claim 1, wherein extracting the plurality of features comprises identifying semantic embeddings.Claim 4
The method of claim 1, further comprising storing flagged content items in a review queue database.Claim 5
The method of claim 1, wherein the predetermined threshold is dynamically adjusted based on historical accuracy metrics.Evidence: Source Code Analysis
ContentModerationService.java, lines 142-156: processSubmission() method receives content via API endpoint and initiates classification pipeline.Evidence: Technical Documentation
Architecture Doc v2.3, Section 4.1: "The feature extraction module utilises a pre-trained transformer model to generate 768-dimensional embeddings."Evidence: Source Code Analysis
ClassifierModel.py, lines 89-102: forward() method implements scoring logic using extracted feature vectors.Evidence: Configuration Files
moderation_config.yaml: threshold_score: 0.85, action_on_exceed: "FLAG_FOR_REVIEW"Evidence: API Response Logs
Sample response showing automatic flagging when score > threshold: {"status": "flagged", "score": 0.91, "review_queue": "priority"}Evidence: Model Architecture
model_spec.json confirms ResNet-50 backbone with custom classification head, consistent with CNN architecture.Evidence: Database Schema
review_queue table schema includes flagged_content_id, timestamp, priority_score columns matching claim requirements.Evidence: Configuration Logic
ThresholdManager.java implements adaptive threshold adjustment based on precision/recall metrics from validation set.Evidence: Embedding Generation
EmbeddingService.java, lines 45-78: generateSemanticEmbeddings() uses sentence-transformers to create 768-dimensional vectors.Evidence: Prior Art Reference
US Patent 9,123,456, Fig. 3: Similar CNN architecture for image classification disclosed in 2015.Claim 1 [preamble]
A computer-implemented method for content moderation comprising:Claim 1.1
receiving, by one or more processors, a content item submitted to an online platform;Claim 1.2
extracting, by the one or more processors, a plurality of features from the content item using a trained machine learning model;Claim 1.3
generating, by the one or more processors, a classification score based on the extracted features;Claim 1.4
comparing the classification score to a predetermined threshold value;Claim 1.5
in response to the classification score exceeding the threshold, automatically flagging the content item for review.Claim 2
The method of claim 1, wherein the trained machine learning model comprises a convolutional neural network.Claim 3
The method of claim 1, wherein extracting the plurality of features comprises identifying semantic embeddings.Claim 4
The method of claim 1, further comprising storing flagged content items in a review queue database.Claim 5
The method of claim 1, wherein the predetermined threshold is dynamically adjusted based on historical accuracy metrics.Evidence: Source Code Analysis
ContentModerationService.java, lines 142-156: processSubmission() method receives content via API endpoint and initiates classification pipeline.Evidence: Technical Documentation
Architecture Doc v2.3, Section 4.1: "The feature extraction module utilises a pre-trained transformer model to generate 768-dimensional embeddings."Evidence: Source Code Analysis
ClassifierModel.py, lines 89-102: forward() method implements scoring logic using extracted feature vectors.Evidence: Configuration Files
moderation_config.yaml: threshold_score: 0.85, action_on_exceed: "FLAG_FOR_REVIEW"Evidence: API Response Logs
Sample response showing automatic flagging when score > threshold: {"status": "flagged", "score": 0.91, "review_queue": "priority"}Evidence: Model Architecture
model_spec.json confirms ResNet-50 backbone with custom classification head, consistent with CNN architecture.Evidence: Database Schema
review_queue table schema includes flagged_content_id, timestamp, priority_score columns matching claim requirements.Evidence: Configuration Logic
ThresholdManager.java implements adaptive threshold adjustment based on precision/recall metrics from validation set.Evidence: Embedding Generation
EmbeddingService.java, lines 45-78: generateSemanticEmbeddings() uses sentence-transformers to create 768-dimensional vectors.Evidence: Prior Art Reference
US Patent 9,123,456, Fig. 3: Similar CNN architecture for image classification disclosed in 2015.We price litigation risk.
Calibrated, reproducible probabilities for litigation assets — cases, judgments, and portfolios. Every prediction timestamped before the outcome resolves.
195MPatent records ingested
926KLitigation cases
5.76BRelationships mapped
340K+Litigants profiled
50+Jurisdictions
Every prediction hashed, tied to a model version and data freeze, and lodged before the outcome can resolve. Same inputs, same model, same freeze — byte-identical result, every time.
Products
A calibrated number, in the format you price in
One probability engine, four renders. Each carries its model version, data freeze, and the calibration record behind the number.
API
Everything above, programmatic. Screens, entity resolution, monitoring webhooks; reversal and portfolio endpoints as they ship. Deterministic: same inputs, same model version, same freeze — byte-identical result, every time.
Get API access→$ curl -X POST https://api.wearemonsters.co.za/v1/screens \
-d '{"patents":["US12574599B2"],"defendant":{"name":"..."}}'
-d '{"patents":["US12574599B2"],"defendant":{"name":"..."}}'
Calibration record
We publish what we got right
01
Data frozen
Inputs locked to a timestamped snapshot.
02
Prediction registered
Hashed and lodged before the outcome can resolve.
03
Outcome resolves
Auto-resolved from the public docket.
04
Published quarterly
Predicted-vs-realised by band and product.
“Any system that only predicts wins is useless. Intelligence is valuable precisely because it tells you when not to proceed.”
The Platform
From a single case to a whole book
The same infrastructure that scores one matter prices an entire portfolio — duration, concentration, and outcome, position by position.
01Assess
02Identify
03Compare
04Execute
05Monitor
Target intelligenceWho to enforce against, and in what order
01Target A
94Settlement: 73%Prior cases: 12
02Target B
81Settlement: 65%Prior cases: 8
03Target C
67Settlement: 48%Prior cases: 3
Strategy comparisonWhich enforcement approach produces the best outcome
Germany first
72%
14 monthsModerate
Multi-jurisdictionRecommended
91%
8 monthsHigher upfront cost
US district court
64%
18 monthsAppeal risk
Outcome predictionWhat will happen if you file, with confidence ranges
Settlement
82%
Trial
12%
Dismissal
6%
Expected duration9–14 mo
ConcentrationModerate
Based on resolved historical cases, judge patterns, and defendant behaviour — with duration and concentration for portfolio marks. Confidence narrows as the case progresses.
How We HelpWhat we
View all services→What we
build
The platform produces the score. Our experts produce the work product — claim charts, prior art, testimony, the signed opinion. The number is neutral; the opinion is ours, and signed.
Why Us
How we work
Case Studies
View all case studies→See what it predicted
Knowledge
Learn from our experience
In-depth articles on patent processes, technical analysis, and practical guidance.
Talk to us about your book.
No pitch deck. No demo. Just a conversation about what you're working on.
Book a Call