HEIMDALL
Advanced Language Analysis / Knowledge Discovery / Content Safety
A 900M-parameter DeBERTa model classifying 56 span-level manipulation techniques, 8 stealth propaganda analyzers exposing bias hidden behind journalistic objectivity, a narrative flow engine mapping 16 manipulation choreography patterns across 6 emotional arc shapes, and a 12-plugin content safety stack from CSAM detection to grooming prevention.
Explore Solutions →The Problem: Manipulation You Can't See
Fact-checkers verify claims. Heimdall detects the psychological machinery behind them. An article can be 100% factually accurate and still manipulate you through selective sourcing, emotional quote stacking, strategic hedging, and carefully sequenced narrative arcs.
Our narrative flow engine identifies 16 distinct manipulation choreography patterns — from “sympathetic hook” openings that lower critical defenses, to fear escalation sequences that build urgency, to call-to-action buildups timed for maximum emotional engagement. These patterns map onto 6 fundamental story shapes documented in computational narrative research.
Our stealth propaganda analyzers go further: measuring source perspective ratios, detecting emotional quote asymmetry via VADER sentiment scoring, identifying word-choice framing through SpaCy dependency parsing, and flagging strategic hedging across 40 doubt-planting patterns. This is manipulation that passes every fact-check — and Heimdall catches it.
Solutions for Every Domain
One platform, tailored for your mission.
Content Analysis
Security & Intelligence
Industry Solutions
Signal Intelligence
Recently shipped capabilities powering the Heimdall pipeline.
Signal-to-Noise Ratio
Measures information quality versus rhetorical noise across 6 sub-metrics. Composite score 0-100 with noise bands. Does not judge truth — measures how much of the content is signal versus noise.
Content Word Ratio
Loaded Language
Claim-Evidence Ratio
Hedge Density
Repetition
Semantic Density
Noise Bands: Low (0-33), Moderate (34-66), High (67-100). User signals include loaded_language, low_evidence, deniability_language, circular_rhetoric, and low_substance.
At a Glance
Powered by local Gemma 4 model — on-premises, no cloud API costs. Generates structured cliff-notes summaries replacing RSS descriptions with distilled intelligence.
Summary
Distilled article summary
W5H
Who / What / When / Where / Why / How
Context
Background and framing
Claims
Key assertions extracted
Sources Cited
Referenced sources listed
Claims extraction enables future fact-checking integrations.
Readability & Writing Level
Consensus reading level computed across multiple algorithms. Writer-versus-audience level gap analysis detects when content is written below its audience's level — a manipulation risk flag indicating deliberate simplification to bypass critical thinking.
5-Pathway Influence Score
Updated from 4 pathways to 5. Final score equals the maximum across all pathways, scaled 0-100.
Overt Propaganda
Density, severity, and diversity of detected propaganda techniques
Stealth Manipulation
Stealth analysis score combined with low evidence and loaded language signals
Emotional Manipulation
Narrative flow score combined with loaded language and sentiment analysis
Balanced Manipulation
All dimensions weighted together plus noise score
Information Quality
NEWNoise score, unsupported claims, repetition, and writing level manipulation risk
Built on Research-Grade AI
Not keyword matching. Not simple sentiment analysis. A 900M-parameter multi-task transformer with five simultaneous classification heads, focal loss for class imbalance, and temperature-calibrated confidence scoring.
DeBERTa-v3-large
900M parameter transformer fine-tuned for manipulation detection. FP16 inference with temperature-calibrated confidence scores.
5-Head Multi-Task Model
Simultaneous token-level technique classification (56 labels), hierarchical category mapping (7 categories), intensity regression, emotion detection (7 classes), and document-level propaganda scoring — all in a single forward pass.
Real-Time Inference API
Dynamic batching (up to 32 requests per GPU pass, 100ms collection window), LRU cache with 10K entries, sub-second latency. Optional LLM explanations via Ollama, OpenAI, Gemini, or Claude.
Plugin Architecture
Modular detection pipeline: 8 stealth propaganda analyzers, 12 content safety plugins, 16 narrative flow patterns, and entertainment analysis — each running independently with async orchestration.
Narrative Flow Engine
Emotional arc classification across 6 fundamental story shapes (Reagan et al. 2016), Freytag dramatic structure mapping, and 16 manipulation choreography patterns including fear escalation, sympathetic hooks, and confirmation bias exploitation.
Content Safety Stack
12 plugins spanning text toxicity (Detoxify RoBERTa), NSFW detection (NudeNet v3), CSAM hashing (PhotoDNA + Thorn), grooming detection (4 sub-analyzers), video moderation (ffmpeg + Whisper + CLIP), and URL scanning — all configurable across 4 audience tiers.