MANA AI MANA Vision : how it works

Card scanner that actually scans cards.

Real AI Vision running locally on an IoT-powered kiosk + the MANA cloud. Rotated detection, region-aware analysis, multi-signal foil chain, illumination-aware confidence. No cloud-grading vendor in the loop. No "AI" without the actual AI.

Try MANA Vision → 🧪 Experimental
SiMANA
99.7%
Identity accuracy
Right card name across our full benchmark suite
95.3%
Exact-print accuracy
Right name AND exact printing, finish, and treatment — the strictest tier
96.6%
Commercially-material accuracy
When we miss the exact variant, the one we pick is within $5 of the true card's value — buyers don't get burned
98.4%
Foil & finish verdict
Live accuracy on foil, etched, surge, silverscroll, and non-foil — illumination-aware
The pipeline
A scan moves through six stages from photo to verified match. No black box — every step is auditable, with per-stage confidence preserved end-to-end.
CaptureMobile photo, any angle, any lighting
DetectAI Vision → multi-region rotated detection
WarpPerspective-correct to a canonical view
MatchAI image fingerprint matcher
VerifyMulti-signal foil chain
OCR fallbackAdvanced OCR on collector strip
What makes it different
Six things that separate a working in-house scanner from "AI pre-grading" marketing.
⚡
Rotated detection
Cards photographed at angles still get clean bounding boxes. Sub-regions stay aligned with the card so the warp stays straight.
🎯
Multi-region analysis
Different parts of every card are analyzed independently — each contributes its own confidence signal. No all-or-nothing verdict.
✨
Foil chain
8 ordered signals: always-foil sets → prerelease → era hard rule → finishes-only printing → ★ on collector → graded slab → classifier → sleeve-glare guard.
💡
Illumination-aware
UV vs white-LED vs ambient light all bias the foil-decision threshold. Verdicts stamp the cutoff so staff can trace which threshold fired.
🛡️
Hostile-trained
Trained on real adversarial scans (cropped, finger-obscured, multi-card frames, glare). Robust by design, not by averaging clean studio photos.
🔍
Inspectable
Every detection saved with bbox quads, foil source attribution, illumination mode. No black box. Staff can audit, correct, retrain.
Where it's deployed
Same model, four contexts. The scanner is a feature, not a product silo.
📱
Customer Scanner
Snap a photo, get an instant match + price. One tap to add to collection.
🤖
IoT-powered Kiosk
In-store scanning station with controlled lighting (UV + white-LED) and fixed camera distance. Built for speed and accuracy.
🏷️
Staff Labeling
Every scan reviewable in our internal labeling console. Manual corrections feed the next training round.
📊
Live Stats
Internal telemetry tracks accuracy, foil source attribution, and confidence drift per release. Continuous calibration.
SiMANA
FYI FROM MANA