Content Moderation API
⚡ Performance KPIs (measured)
| Metric | Brainiall | AWS Rekognition | |
|---|---|---|---|
| End-to-end pipeline p50 latency | 91 ms [source] | 200–400 ms [source] | ★ |
| Binary NSFW p50 | 62 ms (Brainiall quantized+compiled) | ~150 ms | ★ |
| Region detect p50 | 22 ms (NudeNet v3 optimized inference) | ~200 ms (3-level taxonomy) | ★ |
| Throughput per CPU core (sustained) | 8 RPS | Cloud auto-scale | = |
| Cold-start | ~5 s (always-warm container) | <1 s | R |
🎯 Capability matrix
| Metric | Brainiall | AWS Rekognition | |
|---|---|---|---|
| Binary safe/unsafe classification | ✅ Brainiall (98.04% on 80k eval) | ✅ DetectModerationLabels | = |
| Granular region detection (body parts) | ✅ NudeNet 12 classes (breast, genitalia, etc.) | ✅ 3-level taxonomy (Explicit/Suggestive/etc.) | R |
| L1/L2/L3 hierarchy taxonomy | 🟡 Binary + region (no hierarchy yet) | ✅ L1 + L2 + L3 fine-grained labels | R |
| Violence/weapon/drugs detection | ❌ v1 covers NSFW only — v1.1 planned (custom YOLOv11) | ✅ Built-in | R |
| Local CPU deployment (no cloud RTT) | ✅ Latitude bare-metal | ❌ Cloud-only | ★ |
| Open weights you can audit | ✅ Brainiall permissive license + NudeNet MIT | ❌ Proprietary closed | ★ |
| LGPD / GDPR-by-default | ✅ EU/BR datacenter | 🟡 us-east default | ★ |
| Pricing | $0.0008/img (undercut) | $0.001/img | ★ |
📊 Quality benchmarks (literature)
| Metric | Brainiall | AWS Rekognition | |
|---|---|---|---|
| NSFW classification accuracy | Brainiall 98.04% (80k proprietary eval) | Claim 'up to 95% unsafe content flagged' [source] | |
| Region detection coverage | NudeNet 12 body-part classes + face | Hierarchy of ~30 fine labels [source] | |
| False-positive rate (LATAM context) | Higher on Brazilian beach/carnival imagery (cultural skew, fine-tune planned) | Same skew (US-trained) | = |
Pricing
Free
1,000 imgs/month
Get started. No card.
Fast
$0.0008 / image
Brainiall-quantized + NudeNet. p50 91 ms. Undercut Rekognition by 20%.
Pro
$0.005 / image
Adds custom-fine-tuned LATAM model + L1/L2/L3 hierarchy (Q3 2026).
Quickstart (Python)
Request
import base64, httpx
img = base64.b64encode(open("photo.jpg", "rb").read()).decode()
resp = httpx.post(
"https://api.brainiall.com/v1/moderation/analyze/base64",
headers={"Authorization": "Bearer brnl-..."},
json={"image": img, "include": ["binary", "regions"]},
)
print(resp.json())Example response
{
"request_id": "req_b3f9c2…",
"processing_ms": 91,
"is_safe": false,
"binary": {
"label": "unsafe",
"score": 0.987
},
"regions": [
{
"label": "EXPOSED_BREAST_F",
"score": 0.91,
"box": [124, 88, 312, 290]
}
],
"warnings": []
}⚠️ CSAM out of scope
Brainiall's Content Moderation does NOT detect CSAM. For CSAM detection, route uploads through Microsoft PhotoDNA, Thorn Safer, or report directly to the NCMEC CyberTipline (18 U.S.C. § 2258A). Full model card with per-class metrics, known biases, and audit trail commitment: /trust/content-moderation-model-card.
Comparison methodology & disclaimer
Brainiall measurements: latency from Brainiall production infrastructure (Latitude bare-metal · production hardware · CPU), May 2026. Models: Brainiall NSFW (quantized+torch.compile per Phase 1.5) + NudeNet v3 (MIT). Full report: Phase 1.5 Eval Report.
AWS Rekognition data: derived from AWS public documentation (May 2026). AWS does not publish formal accuracy benchmarks; their published claim of "up to 95% unsafe content flagged" is cited as-is. Latency ranges include network round-trip from us-east-1.
Important notes:
- Brainiall S7 v1 covers NSFW classification + body-part regions. Violence, weapons, drugs, hate symbols are NOT in v1 — planned for v1.1 (custom YOLOv11 trained on Roboflow public weapons + violence datasets). Rekognition has these built-in via taxonomy, so for those categories AWS currently has clear coverage edge.
- L1/L2/L3 hierarchy taxonomy: AWS provides this out of the box, we currently provide flat binary + region — hierarchy is on roadmap Q3 2026.
- LATAM context: both systems trained predominantly on Western datasets; false-positive rate on Brazilian carnival/beach/fashion imagery is comparable. Brainiall plans local fine-tune in Q3 2026.
- Methodologies and datasets may differ — direct head-to-head benchmarks scheduled for Q3 2026.
- Trademarks: Amazon Web Services and Rekognition are trademarks of Amazon.com, Inc. This page is provided for informational comparison purposes and is not endorsed by or affiliated with AWS.
Last reviewed: May 2026. Sources cited inline; we update when new data is available.