AI camera box tiers and specs

Every Prism site starts with the Prism Box — the "node" at your gym. It records every camera, gives you live view in the app, runs the gate lines and spots possible incidents with a small AI model on the box. Higher tiers add bigger AI models, either on Prism's own model server or on a more powerful box at your site.

A side-by-side version of this page is at prismgym.com/ai-camera-box.

Tier 1 — Essentials (available now)

Hardware: the Prism Box — an Intel N100/N150 mini PC with 16 GB RAM, a 512 GB SSD and Intel graphics.

It runs:

• 24/7 recording, and live view in the app
• Gate lines drawn per lane: unscanned entry, tailgating, and wrong-way entry through the exit
• Overhead people counting, and people-inside-now on the dashboard
• After-hours presence alerts
• Telegram alerts
• A per-site setting for every rule

It uses a compact detector on the box. That is good for straightforward gates; busy scenes produce some false alarms, which staff review in the app. It supports about 2–4 cameras.

Tier 2 — Verified (pilot, rolling out)

Hardware: the same Prism Box, plus Prism's own model server (a monthly service).

• Every gate incident clip is re-counted by a much larger, state-of-the-art detection model (RF-DETR class) with a proper multi-person tracker on Prism's server: who crossed which lane and which way, checked against the door scans — for far fewer false alarms.
• Optional AI review of incidents — reception left unattended, a customer waiting, staff on the phone, incident summaries — uses still images with every face blurred on Prism's own server before review. Never live video. See AI modules.

Tier 3 — Pro on-site (on request, 4+ cameras)

Hardware: a more powerful on-site AI box with Intel Arc graphics or an NVIDIA graphics card.

It runs the large model live, at your site:

• Accurate counting and tracking through crowds
• The large face-recognition model
• Matching people across cameras

The AI work happens in the building — nothing leaves it.

Specs at a glance

• Availability — Essentials: now · Verified: pilot · Pro on-site: on request
• On-site hardware — Essentials and Verified: Prism Box (Intel N100/N150, 16 GB RAM, 512 GB SSD, Intel graphics) · Pro on-site: Intel Arc or NVIDIA graphics
• Where the AI runs — Essentials: on the box · Verified: on the box, plus Prism's model server · Pro on-site: on the Pro box, live
• Cameras — Essentials and Verified: about 2–4 · Pro on-site: 4 or more
• Large face-recognition model, cross-camera matching — Pro on-site only

Why the big model doesn't run live on the Prism Box

On the Prism Box, the compact detector takes about 15 ms per check; the state-of-the-art model takes about 186 ms on the same box — too slow to run live there. That is why it runs on Prism's model server (Verified) or on a Pro box (Pro on-site).

In a test on real gate footage, the larger model found the person in 15 of 15 frames at the turnstile, against 12 of 15 for the compact model — and 14 of 15 against 9 of 15 on the overhead counter.

Where your video stays

Video and recordings stay on the box at your site and on Prism's own infrastructure. Any AI review by an outside model only ever sees face-blurred still images.