Body recognition
Tells the difference between staff (enrolled at install) and unfamiliar visitors. After-hours unfamiliar-face alerts go to your phone in <3 seconds.
What the AI actually does
A recording-only system makes you scrub through 24 hours of footage to find the 30 seconds that mattered. The AI worker watches it for you and produces a plain-English daily digest plus immediate alerts when something needs your attention.
License-plate flagging
Allow / watch / block lists. Auto-alert when a flagged plate enters. Useful for repeat shoplift cases, persistent loiterers, and chronic parking issues.
Motion-first gate
Quiet rooms run zero LLM calls. The classifier only fires when a cheap pixel-diff thumbnail says something changed. Saves GPU cycles + electricity.
Tampering detection
Camera covered, moved, or losing signal? You get a phone alert. Most break-ins start with someone covering or pivoting a camera; the AI catches that move.
Daily plain-English digest
Emailed every morning at 7 a.m. (you pick the time). "12 vehicles arrived after-hours, 3 unfamiliar visitors, 1 plate on watchlist." Read it with your coffee.
Event search in plain English
Type "show me anyone who entered the front lobby alone after 9 p.m. last week" and the system returns the matching clips. No code, no SQL, no scrubbing.
The privacy model is the point
Most "AI camera" systems send your footage to a third-party cloud where it is processed and stored. The whole reason this tier exists is to flip that arrangement.
What stays on your premises
- All raw footage (recorded to your local NVR).
- All face / body embeddings (stored in a local SQLite DB).
- All license-plate reads (local DB, exportable to CSV).
- All AI inference (runs on the local RTX GPU).
- The daily digest itself (emailed from your network).
- Event audit log (every classification + timestamp).
What is NEVER uploaded
- No raw footage to a third-party cloud.
- No face embeddings to a third-party model provider.
- No plate reads to a Flock-style aggregation network.
- No anonymized telemetry to "improve our models."
- No staff faces to anyone, ever.
- No customer faces to anyone, ever.
The optional exception: off-site recording-INDEX backup (just the metadata, not the footage) goes to a destination YOU control — your own cloud bucket or your own off-site NAS. Opt-in only. Even this can be turned off if you want fully air-gapped operation.
What it runs on
- Compute: a small server (mini-tower or 1U rack) with an RTX 5060 8GB or RTX 4060 GPU. Sized so the AI vision worker has dedicated VRAM without competing with anything else on your network.
- Power: ~50W idle, ~120W under classification load. Roughly $9/month on most utility rates running 24/7.
- Models: face recognition via InsightFace (open-source), license-plate via PaddleOCR (open-source). All weights run locally; nothing phones home.
- Software: the same TCG vision-guard worker that runs on the operator's own fabric, hardened for production. NSSM-managed Windows service. Auto-restarts on failure.
- Storage: uses your existing NVR for raw footage. Adds a small (~100GB) local SQLite for embeddings + event log. No additional storage requirement.
Pricing notes
How AI Integration is priced: AI Integration is an add-on quote on top of any Basic, Advanced, or Custom install. Most adds land between $1,800 and $4,200 depending on whether you need a new GPU machine or can repurpose an existing PC, plus the per-camera enrollment time.
- One-time hardware — $850-$1,400 for a mini-tower with RTX 4060/5060, 32GB RAM, 1TB NVMe.
- Software install + tuning — flat $750 covering install, motion-zone tuning, staff-face enrollment (up to 25 staff).
- Per-camera tuning — $40 per additional camera beyond the first 4 (each needs zone tuning to keep false-positives low).
- No monthly fees. No subscription. The system is yours to run forever.
- Verified nonprofit code TCG26FREE applies to install + tuning; hardware not included.
Common questions about AI Integration
Does AI Integration require new cameras?
No. It works with any IP cameras you already have, as long as they expose RTSP or ONVIF. The AI worker reads from your NVR's RTSP streams; no new cameras needed.
What about HIPAA/privacy if I have customers in the cameras?
The system is designed around privacy from day one — face embeddings stay local, no third-party uploads. For HIPAA-sensitive sites (dental, medical), the install includes a zone-exclusion config so exam-room interiors are never recorded or analyzed. See the Healthcare industry page for the full HIPAA-aware setup.
Can I tell the AI to ignore certain things?
Yes. Per-camera motion zones (ignore the highway in the corner of frame), per-classifier sensitivity (loud false-positives turned down), per-time-of-day rules (only alert on unfamiliar faces between 7 p.m. and 6 a.m.).
What if the AI gets it wrong?
Every classification is logged with the source frame. If the AI flagged a delivery driver as "unfamiliar," you click "false positive" in the daily digest and that person becomes part of the rolling allow-list. The system tunes itself across the first 30 days; after that false-positive rates settle below 2%.
If I cancel later, can I take the system with me?
Yes. Day-one handoff includes admin credentials for the GPU server, the AI worker config, and the SQLite DB. Any other competent IT shop can pick up where I left off — nothing is licensed or DRM'd.