AI video intelligence platform

Your cameras already see everything. Now they understand it.

MIT Eyes turns the CCTV you already own into a tireless AI analyst — detecting people, vehicles, animals, fire, smoke and weapons in real time, recognising faces and plates, making every second of footage searchable, and emailing the right person the moment something matters. All of it self-hosted, on your hardware.

DETECTION AREAperson 97%plate B 4173 XKFIRE 92%rule fired · email sent ✓LIVE · cam-04 · schema “Perimeter”3 detections · indexed
24/7
attention — an AI never gets tired or looks away
< 1 s
from something happening to an alert firing
face engines on every detection — nothing slips by
0 code
automations — draw the rule, the emails send themselves
0
cameras replaced — plugs into your existing CCTV
what you get

Everything the job needs, in one system

Connect any camera

RTSP, ONVIF, MJPEG, HLS, RTMP and WebRTC — IP cameras, NVRs, drones or phones stream straight in. No proprietary hardware, ever.

Find the cameras for me

The built-in network scanner sweeps a subnet, lists every live device and infers which ones are cameras — so onboarding a site you inherited doesn't start with a spreadsheet.

Vision schemas

Author what to detect and recognise once — “Perimeter, high accuracy”, “Lobby, faces only” — then assign that profile to any number of cameras. Retune the whole estate by editing one schema.

Detection areas

Draw the part of the frame that matters and the pipeline ignores the rest. The street behind your gate stops waking the AI — and stops generating noise.

Real-time detection

People, vehicles and animals detected live with motion-gated AI, so compute is spent on moments that matter — not on empty hallways.

Hazard alerts

Fire, smoke, sparks, weapons and knives trigger instant on-screen and email alerts — seconds matter, and MIT Eyes reacts in under one.

Dual-engine face recognition

Every face is embedded by two independent recognisers (ArcFace + AdaFace). Hard angles, low light, ageing footage — one engine catches what the other misses.

Licence-plate reading

ANPR with fuzzy matching that forgives single-character OCR misreads, plus a dedicated tracking mode that votes every frame of a car's pass into one plate.

Workflow builder

Drag camera, subject, time, frequency and email onto a canvas and you have an alert rule. “Unknown face at the back door after 20:00, at most one email an hour” takes a minute to build and no developer.

Operations dashboard

A security-operations command view: system posture at a glance, live threat feed, camera health, activity per hour and a watchlist of what still needs a human.

Model tiers you control

Pick the accuracy/speed tier per function — detection, faces, plates, hazards. Weights download in the background and the switch only flips once they're ready, so nothing ever goes dark.

Archive mining

Point MIT Eyes at years of recorded footage — local disks, network shares or cloud buckets — and retire the originals as a searchable database of detections.

Search, don't scrub

“Show me every white van since Tuesday” is a two-second query, not a two-day shift of watching timelines.

Sites & locations

Model the real world — sites, buildings, camera locations — so every detection carries where it happened and multi-site estates stay legible.

Reports & capacity planning

Traffic patterns, occupancy trends and hardware sizing built in — evidence for decisions, not gut feel.

Self-hosted by design

Runs entirely on your own servers, with or without a GPU. No cloud account, no per-camera subscription, no footage leaving the building.

how it works

From camera to outcome in five steps

1Connectany camera, any protocol2Focusschema + detection area3Understandfaces · plates · hazards4Automaterules that email people5Searchindexed forever
why teams switch

Traditional CCTV / VMS vs MIT Eyes

The traditional wayMIT Eyes
Finding an incidentHours of scrubbing timelines, hoping you don't blink at the wrong momentType a query — every person, plate and event is indexed and found in seconds
When you learn about itAfter the damage is done, during footage reviewThe moment it happens — real-time alerts for hazards, faces and plates
Who tells youSomeone has to be watching the video wall — and be the right someoneRules you draw once email the right person, with the evidence attached
Who's watchingA human operator whose attention collapses after ~20 minutesAI on every frame of every camera, around the clock, with equal focus
Retuning a cameraA vendor visit, a per-device config, and a change windowEdit one vision schema — every camera assigned to it follows within seconds
False alarmsEvery passing car on the road behind the fenceDetection areas confine the AI to the ground you actually own
StorageTerabytes of raw 24/7 footage nobody will ever watchMeaningful, indexed events with photo evidence — a fraction of the footprint
HardwareProprietary NVRs, licensed channels, forklift upgradesRuns on your existing cameras and commodity servers
Your dataLocked in a vendor cloud, per-camera subscription feesFully self-hosted on your premises — your footage never leaves the building
the difference, in numbers

Measured against the old way

Time to find one incident in a week of footage
traditional~4 hours of scrubbing
MIT Eyesseconds — indexed search
Video your team can actually monitor live
traditionala handful of screens, 20-min focus
MIT Eyesevery camera, every frame, 24/7
Effort to retune detection across 40 cameras
traditional40 devices, one at a time
MIT Eyesone schema edit
Storage kept per month
traditionalterabytes of raw 24/7 footage
MIT Eyesindexed events with evidence
competitive edge

The advantages competitors can't copy overnight

  1. 01

    Two face engines, one enrolment

    Every face is stored with both ArcFace and AdaFace embeddings. Switch engines per camera at any time — no re-enrolment, no lost history. Almost nobody else in the market does this.

  2. 02

    Tune an estate, not a camera

    Vision schemas turn detection settings into a reusable profile. Change one schema and every camera carrying it retunes itself — the difference between managing 8 cameras and managing 800.

  3. 03

    Automation without a developer

    The workflow builder turns “tell facilities when an unknown face appears at the loading bay after hours” into a five-node diagram. Competitors ship an API and a quote for professional services.

  4. 04

    Compute spent on meaning

    Motion gates and drawn detection areas discard irrelevant frames before the heavy AI runs, so one modest GPU covers a whole site — competitors quote a server rack.

  5. 05

    100% on-premises

    Faces and plates are sensitive data. MIT Eyes keeps them on your hardware with zero cloud dependency and zero recurring per-camera fees.

  6. 06

    Your archive becomes an asset

    Years of DVR recordings stop being dead weight: mine them into searchable detections, then reclaim the storage.

  7. 07

    A platform, not a dead end

    The same cameras and the same install power MIT Parking, MIT Attendance and MIT Traffic — each added module multiplies the return on the first.

the platform play

Install once. Monetise four ways.

Your cameras+ one on-prem serverMIT EyesAI video intelligenceyou are hereMIT ParkingANPR parking revenueMIT Attendanceworkforce attendanceMIT Trafficroad counts & enforcement

Competitors sell four systems: four installs, four vendors, four invoices. MIT sells one platform where every module reuses the cameras, the AI and the data of the others — so each solution you add costs a fraction and returns in full.

  • One enrolment, everywhere — a face enrolled for security also powers attendance.
  • One plate database — investigations and parking sessions share the same ANPR.
  • One set of rules — the same workflow builder emails on a hazard, a watchlist plate, a no-show or an over-speed.
  • Software, not hardware — the next module is a licence, not a construction project.

Every day on traditional CCTV is footage you can't search and incidents you find too late.

See MIT Eyes running on your own cameras — a pilot takes a day, not a quarter.