counting · live
tracking · forklifts
produce · sorting
machine · state
proximity · safety
collision · alert
fall · alert
packing
heatmap · dwell
Darkfield · UK private beta · 2026

Watch
what your
cameras
already see.

AI vision in production — across waste, food, manufacturing, and forecourts.

see the platform → explore use cases
What it actually does

You give
us a feed.
We tell you
what's on it.

Why Darkfield

Computer vision
isn't new.
We are.

Hundreds of CV companies count people, track forklifts, and draw zones on your feed. We do all of that — and the parts traditional vision can't.

01

Auto-training

No manual labelling. No engineer-hours of sync. The pipeline trains itself on your scene — costs collapse from a five-figure setup to a marginal one.

  • no manual labels
  • no engineer-hours
  • cost collapses
02

Context-aware

Spatial and temporal reasoning across cameras and shifts. Sees the line stopped — and tells you which station, how long, and why. Distinguishes intent.

  • spatial + temporal
  • across cameras + shifts
  • understands why
03

Vision AI layer

A VLM reads each clip the way a human would, reads your SOPs, calls the APIs you connect, and filters the noise. Only what's worth interrupting someone for.

  • reads your SOPs
  • calls your APIs
  • filters the noise
Tier-1 line system £40k–£150k 6–12 weeks of integration
→
Darkfield £200 / camera / year 48 hours to first alert
Use cases shipping · 2026

Ten things
we already do well.

Each comes with the alert your team will actually act on, and a one-page weekly summary they will actually read.

use case / 01 Waste reduction
Fewer collections, cleaner streams, contamination caught the moment it lands — not at the audit. read →
use case / 02 PPE compliance
Hi-vis, gloves, hairnets, masks — checked on every person, every camera, every minute. Per-area policy. read →
use case / 03 SOP adherence
Cleaning cycles, assembly steps, torque sequences, machine resets, hand-washes — was each one done, in the right order, for long enough? read →
use case / 04 Bottleneck attribution
When a line stops, the supervisor's phone rings within seconds — which station, how long, and why. read →
use case / 05 Quality at end of line
Visual QA on the conveyor — defects, miscounts, and packaging faults caught before they ship. read →
use case / 06 Safety & incident
Falls, collisions, restricted-area breaches — flagged in seconds, indexed for the post-incident report. read →
use case / 07 Security
Not just a person on camera — a person doing something. A VLM reads the clip and tells you whether to care. read →
use case / 08 Visit analytics
Footfall, dwell time, and route patterns from CCTV alone — no beacons, no apps, no data-protection grief. read →
use case / 09 Forecourt analytics
Vehicle classification, ICEing prevention, charging-bay overstays, plates checked against DVLA in seconds. read →
use case / 10 Object counting
Items entering or leaving a zone, counted by category, in real time. Throughput rates, end-of-shift totals, no manual tally. read →
explore all use cases →
Three things · Every site · Every day

We watch.
We alert.
We help you act.

We work with the cameras you already have. Where a camera doesn't fit the job — wrong angle, wrong shutter, wrong focal length — we ship and install one that does. Either way, you don't manage hardware. You just receive the answers.

01Watch

Plug in. We watch.

We ship you a pre-configured network gateway and an AI vision-box. You plug the gateway into your local network switch, then plug the vision-box into the gateway. That's the whole install — no IT team needed, no software, no firewall changes. The gateway only reaches the cameras you nominated; the rest of your network stays unreachable.

  • no IT skills needed
  • cameras only — no wider network access
  • works with your existing CCTV
02Tell you

Alerts when it matters.

When something needs a person, we tell them. SMS to the supervisor, voice call to the on-call manager, email to compliance, a webhook into your own stack. No app to log into. No noise — only what's worth interrupting someone for.

  • SMS · email · voice call · webhook
  • thresholds tuned to your site
  • acknowledged in one tap
03Improve

Reports that change things.

Weekly summaries showing where time, money, and effort go. Made for the person who has to act on them — one decision per page, no charts for charts' sake, no jargon. Raw data is one click away if you want it.

  • live / shift / monthly
  • one decision per page
  • raw data exportable as CSV
/ frequently asked

Seven questions
partners ask first.

Practical answers to the practical concerns. The longer ones live in the technical brief.

001 What do I need to get started? +
Two pre-configured devices ship with your contract: a Darkfield network gateway and an AI vision-box. You plug the gateway into a spare port on your local network switch, then plug the vision-box into the gateway — total install time under fifteen minutes. No IT team, no port-forwarding, no firewall changes, no software on your machines. Anyone on your team can do it. The gateway is scoped to the cameras you nominated and nothing else on your network; the vision-box runs perception on-site so raw video never leaves your firewall by default.
002 How do I know if my camera placement is good enough? +
The AI answers this on day one. Once the gateway is plugged in, the AI pulls a frame from each camera you nominated, runs quality checks against your prompt, and either confirms the camera is sufficient or tells you exactly where to put a different one. If the existing camera isn't enough, we install one — typical install is two hours on-site, minimal disruption. We've onboarded automotive workshops, food production lines, forecourts, and warehouse yards.
003 Where does my video data actually go? +
By default, raw frames stay on your network. The AI vision-box runs perception on-site, behind your firewall — only events, anonymised detections, and short evidence clips reach Darkfield's cloud. For sites where running an edge box isn't practical, a fully cloud-hosted option streams RTSP direct to Darkfield's UK ingest. For defence and critical-infrastructure deployments, an air-gapped option is available. UK GDPR-compliant by default — anonymisation is applied to every frame in RAM before any clip is persisted.
004 My use case isn't a typical vision task. Can it still work? +
Probably. The model is open-vocabulary and type-aware — you don't pick from a class list, you describe what you want in plain language. If you can describe it in plain English, we can usually detect it. For events that need reasoning — "a bottleneck happened at station B", "that person is acting suspiciously" — the Vision AI layer reviews the candidate clip the way a human would and only flags it when the description actually applies.
005 What happens when the model is wrong? +
The AI is its own QA. It samples its own output on a rolling window and scores it against the data table you agreed at onboarding. When precision drops below the tolerance you agreed, the system fires off a retrain automatically — per camera, on the actual scene — and validates the new weights before they go live. You see every change in the audit log; you can pause or override at any time.
006 How long until I have something running? +
Day 0 — the gateway and vision-box arrive pre-configured for your site. Your IT team plugs them into the local network switch (under fifteen minutes). Day 1 — the model is running against your stream with a draft schema. Day 14 — alerts wired to your channels of choice (email, SMS, voice). Week 6 — exit the hands-on onboarding window. The sharper accuracy gains happen in the first 48 hours of operation, automatically.
007 Can I run the whole thing without sending video to the cloud? +
By default, you already are. The vision-box keeps raw video on your side of the firewall — only events and short evidence clips reach our cloud. For full air-gap (no outbound to Darkfield at all), both AI components can run on partner hardware with signed weight updates shipped on disk on a quarterly cadence — required for some defence and critical-infrastructure deployments. See the technical brief for the topology comparison.
UK private beta · 2026

Ready to see
what your cameras
already see?

We're onboarding a small number of partners through the rest of 2026. Send us a feed; we'll tell you within a week what's worth running on it.

see the platform
£200 / camera / year two-week guarantee UK only · 2026