Heimdall SafeSpace
12 Safety Plugins. 4 Moderation Tiers. One API.
The complete content safety stack: 12 moderation plugins spanning text (Detoxify RoBERTa toxicity, ML profanity with adversarial robustness, identity-attack hate speech), images (NudeNet v3 NSFW, CLIP zero-shot violence, SHA-256 + PhotoDNA + Thorn CSAM), video (ffmpeg keyframe + Whisper transcription + EasyOCR + CLIP), conversation (4-analyzer grooming detection), and URL scanning (Google Safe Browsing). All configurable across 4 audience tiers: COPPA (zero-tolerance), Teen, General, and Relaxed. Built on propaganda-grade AI.
Challenges You Face
CSAM Legal Compliance
18 USC 2258A requires detection and NCMEC reporting — our pipeline: SHA-256 hash → PhotoDNA → Thorn Safer → BLOCK_AND_REPORT
Grooming Detection
4 sub-analyzers: IntimacyAnalyzer (escalation), SecrecyDetector (hide from parents), AgeProbeDetector (age questions), StageClassifier — always flags for human review
Community Toxicity
Detoxify RoBERTa model: toxicity, severe_toxicity, obscene, threat, insult, identity_attack, sexual_explicit — 93.74% AUC
Content Volume
Dynamic batching: up to 32 texts per GPU pass. Video pipeline: ffmpeg keyframe extraction every 5 seconds, max 30 frames per video
False Positive Fatigue
ML profanity classifier with fuzzy matching resists l33tspeak and character substitution — not crude keyword lists
Multi-Audience Platforms
COPPA zone: toxicity BLOCK at 0.2, profanity at 0.1. Teen zone: toxicity BLOCK at 0.7, profanity FILTER at 0.5. Same platform, different rules.
Key Features
12-Plugin Safety Stack
Text: toxicity, profanity, hate speech, URL scanning. Image: NSFW, violence, CSAM. Video: keyframe + audio + OCR. Chat: grooming. Plus audit + analytics + rating.
4 Moderation Tiers
COPPA (zero-tolerance), Teen (blocks explicit), General (blocks heavy), Relaxed (flags only) — calibrated thresholds per safety category
CSAM Detection Pipeline
SHA-256 hash → PhotoDNA API → Thorn Safer API — hash-only (never stores images). BLOCK_AND_REPORT action with NCMEC-compliant data.
Grooming Detection
IntimacyAnalyzer + SecrecyDetector + AgeProbeDetector + StageClassifier. Combined risk boosted 0.2 when 2+ signals. Always FLAG for human review.
Video Moderation Pipeline
ffmpeg keyframe extraction (5s intervals, max 30 frames) → NudeNet + CLIP per frame → EasyOCR text → Whisper audio → text safety analysis
NSFW Image Detection
NudeNet v3: EXPOSED_BREAST_F, EXPOSED_GENITALIA_F/M, EXPOSED_BUTTOCKS, EXPOSED_ANUS. Threshold: 0.5. Plus CLIP zero-shot violence (threshold: 0.6).
Hate Speech Detection
Identity attack detection with category classification across 13+ categories — implicit and coded hate included
URL Safety Scanning
Regex extraction → local domain blocklist (instant) → Google Safe Browsing API: MALWARE, SOCIAL_ENGINEERING, UNWANTED_SOFTWARE, POTENTIALLY_HARMFUL_APPLICATION
Techniques We Detect
56 span-level techniques catch recruitment narratives, extremist framing, and ideological manipulation
Coordinated_Messaging and Astroturfing detection identifies organized harassment and inauthentic behavior
4 specialized sub-analyzers detect intimacy escalation, secrecy requests, age probing, and grooming stage progression
Pricing
Platform
- ✓100K MAU
- ✓Full text + image
- ✓CSAM pipeline
- ✓Grooming detection
- ✓All 4 tiers
Enterprise
- ✓1M MAU
- ✓Full 12-plugin stack
- ✓Video moderation
- ✓URL scanning
- ✓API
- ✓Audit logging
12 Plugins. 4 Tiers. Sub-Second Latency.
Schedule a demo or start your free trial today.