Key Partners
- NVIDIA (GPU/CUDA)
- InsightFace / DeepFace
- AWS / GCP · Cloudflare
- Moodle, Canvas, Blackboard
- System integrators · NAAC/ABET consultants
Replacing physical biometric hardware with a BYOD camera model — 3-second AI attendance, fraud-proof anti-spoofing, and bank-grade biometric security.
A systemic operational crisis across higher education and K-12 institutions worldwide.
Manual roll-calls consume 10–15 minutes per lecture — 180 hours lost per classroom per year. Institutions pay faculty salaries for name-calling, not teaching.
Students mark proxies for absent peers. Fingerprint readers fall to silicone molds; RFID cards are shared. Zero empirical auditability of physical presence.
Physical scanners cost $30,000–$100,000+ per campus, plus cabling, queue bottlenecks, hygiene risks, and total failure during power/network outages.
Parents learn of chronic absenteeism only at semester end. Faculty manually transcribe paper sheets into legacy ERPs — slow, error-prone, and opaque.
Any smartphone, tablet, or laptop camera becomes an automated multi-student attendance scanner — physical hardware eliminated entirely.
Runs on existing faculty smartphones & webcams. Zero capital expenditure.
Processes 30–60 students in one live WebRTC sweep or a 15-second video.
95.3% ArcFace accuracy + MiniFASNet anti-spoofing blocks photos, screens & deepfakes.
AES-256-GCM envelope encryption for all biometric vectors. GDPR right-to-erasure built in.
Instant dashboards & alerts for Students, Parents, Faculty, and Directors.
HMAC-signed integrity packs + campus geofence check-in keep working when networks fail.
From a faculty smartphone camera to real-time dashboards — in seconds.
Production splits into isolated web and worker tiers — heavy neural VRAM never touches the web process.
Startup preflight verifies broker connectivity, state sync, shared media, and worker manifest leases before public traffic. Dead worker → /health/workers fails and uploads are refused as typed-retryable.
Engineered for dense classrooms — variable lighting, motion blur, and 35px back-row faces.
Each frame match contributes weight W = Quality × Similarity. Students transition SEEN_ONCE → VERIFIED at threshold — no single-frame false positives.
Top candidate must beat second-best by Δ ≥ 0.04, preventing ambiguous identity assignment in lookalike crowds.
8 stored vectors per student, selected to maximize pairwise cosine distance — distinct poses, lighting, expressions for maximum recall.
Printed photos, screen replays, and deepfakes are blocked before embeddings are ever extracted.
Neural network detects structural screen reflections, bezel edges, and print artifacts. Spoof probability gate: < 0.35.
Local Binary Pattern histograms expose flat print/screen texture. Deliberately bypassed on back-row crops < 50px to avoid false lockouts.
Detects high-frequency spatial interference patterns characteristic of digital screen replay attacks.
Empirical benchmark, 2026-07-28: real classroom footage, 16–26 simultaneous faces per frame @ 3840×2160. Median genuine similarity 0.781 vs median impostor 0.101.
Capture resolution directly dictates whether back-row students are detected in large lecture halls.
| Capture Resolution | Median Face Size | Faces Below Floor (<40px) | Frame Time | 10-Frame Sweep |
|---|---|---|---|---|
| 1280×720 (720p) | 32 px | 71.9% missed | 310 ms | 3.1 s |
| 1920×1080 (1080p) | 47 px | 30.8% missed | 744 ms | 7.4 s |
| 2560×1440 (2.7K — Default) | 63 px | 3.5% — Optimal | 1071 ms | 10.7 s |
| 3840×2160 (4K UHD) | 95 px | 0.0% — Complete | 1233 ms | 12.3 s |
Key takeaway: Moving from 1080p to 2.7K cuts missed back-row detections from 30.8% to 3.5% — with negligible processing overhead. Async video path samples up to 300 frames at stride 18 (90s of footage).
Biometric data is GDPR Article 9 high-sensitivity PII — raw vectors are never stored in plaintext.
AAD binds college_id + student_id + model_version into the GCM tag. Copying an embedding between rows corrupts the tag → SecurityError on decrypt.
POST /api/students/<id>/erase permanently deletes encodings and crop files, and anonymizes audit logs while preserving aggregate statistics.
Shared-schema tenancy anchored by college_id — enforced at both the database kernel and ORM layer.
SET LOCAL app.current_college = '…uuid…';
RLS policies on students, teachers, classrooms, attendances, encodings & parents reject cross-tenant rows at the database kernel. SQLite dev fallback uses a do_orm_execute interceptor injecting tenant filters.
Cryptographically signed offline logging for campuses with poor connectivity.
Colleges define multi-point polygons; check-ins validated via ray-casting containment + face liveness.
Packs valid only for the teacher's scheduled timetable blocks on that calendar day.
/api/offline/sync/manual validates signatures, rejects duplicates, dual-writes to session + attendance tables.
Production-grade, enterprise-proven components across every layer.
22 tables via Alembic migrations · RESTful blueprints with strict role gating.
| Table | Key Fields |
|---|---|
| colleges | geofence_polygon_json, code, is_active |
| teachers | school, department, totp_secret, mfa_enabled |
| students | roll_number, consent_granted, class_id |
| student_encodings | AES-256-GCM ciphertext, model_version |
| attendances | status, confidence_score, verification_method · unique (student, class, date, period) |
| biometric_runs | status, frames, faces_detected, students_verified |
| Endpoint | Access | Function |
|---|---|---|
| POST /api/enroll | Teacher | Register + capture biometrics |
| POST /api/recognize | Teacher | Live sweep (≤2500ms budget) |
| POST /api/attendance_video | Teacher | Async video → task_id |
| POST /api/report_false | Student | Dispute false marking |
| POST /api/students/<id>/erase | Director | GDPR DSAR erasure |
| GET /health/workers | Edge | Worker lease + GPU queue |
| POST /api/campus/checkin | Teacher | Geofence + liveness grant |
Mobile-native interfaces that faculty actually enjoy using.
Parser-blocking inline script sets data-theme before first paint. All tokens via CSS variables — zero hardcoded hex.
Stale-while-revalidate caching for static assets; /api/* and auth routes bypass cache entirely for data security.
The complete commercial architecture on one canvas.
Proprietary SCRFD + ArcFace + MiniFASNet pipeline · AES-256 vault IP · GPU infrastructure · real-classroom benchmark datasets · engineering talent
Engineering 45% · Cloud GPU 30% · Sales & marketing 15% · Security/compliance 10% — 75–80% gross margin
Per-student SaaS (70%) · Implementation fees (15%) · Add-ons & gateways (10%) · Custom integrations (5%)
Four institutional tiers — one decision-maker with budget authority.
Universities, engineering, medical & management institutes with 2,000–50,000+ students.
Strict safety monitoring, mandatory daily records, instant parental communication.
Attendance tied to licensing compliance and government grants.
Mandatory compliance training, certifications, internal workshops.
Directors, Chancellors, Deans, CIOs. Driven by accreditation (NAAC, ABET, NIRF), cost reduction, and fraud elimination.
Faculty need ultra-fast, non-disruptive tools. Students need transparent, dispute-free records.
Parents demand real-time visibility into safety and academic discipline — a powerful retention lever.
Packaged for every institution size — from small academies to multi-campus university systems.
Predictable per-student SaaS — Opex, not Capex — exactly how institutions prefer to pay.
Standard $3.00 · Pro $4.50 · Enterprise $6.00 per student / year
$2,000–$10,000 per campus — roster migration, LMS API integration
WhatsApp/SMS volume packs · on-prem GPU edge licensing ($1,500/server/yr)
Bespoke connectors for proprietary university systems
The BYOD model pays for itself in under 30 days.
| Metric | Legacy Hardware Biometrics | Automatic Attendance BYOD |
|---|---|---|
| Initial Hardware Capex | $45,000 (90 readers @ $500) | $0.00 — faculty smartphones |
| Annual Maintenance Opex | $9,000 (repairs, cabling) | $0.00 — managed cloud |
| Software License | $25,000 / yr ($5/student) | $22,500 / yr ($4.50 Pro) |
| Lost Instruction Time | 15 min/lecture (~$120,000 lost value) | 0 minutes — restored to teaching |
| Buddy Punching Risk | High — card sharing, silicone molds | Zero — AI face + MiniFASNet PAD |
| Payback Period | N/A — continuous hardware loss | < 30 days — immediate net positive |
Optimized async batching makes each GPU worker absurdly efficient.
AI/ML biometric research, backend architects, frontend/PWA engineers.
AWS g4dn/g5 instances, managed PostgreSQL + Redis, Cloudflare egress.
University pilot grants, EdTech conferences, sales commissions.
Penetration testing, GDPR biometric audits, IP & patent filings.
A defensible AI platform with proven performance and immediate unit economics.
Global EdTech — institutional ERPs are the largest growing expenditure segment.
Proprietary multi-stage pipeline (SCRFD + 512D ArcFace + MiniFASNet + voting), AES-256 vault, offline HMAC engine.
Validated on dense classrooms (16–26 faces/frame): 95.3% accuracy, 1.38% wrong-accept rate.
<30-day customer payback, 75%+ gross margins, cost-per-student trending to $0.40/year.
Pilot deployments · University partnerships · Seed investment conversations