Onboarding
You describe the operation, connect a stream, write a prompt. THINK inspects the camera, declares a schema, and proposes the pipeline.
precision
baseline · 0.84
UK private beta · 2026
AI that watches
AI that watches
your cameras.
Tells you what's
happening.
No engineer. No class lists. No labelling. Describe what to track in plain English, point it at your CCTV, and start receiving the answers.
/ during the private beta
Our team handles
Our team handles
everything for you.
We ship the kit. We connect to the cameras you nominate. We dial in the detections and stand up the alerts. You only see the platform once it's already performing to spec.
- ~15 min install effort · no IT skills required
- 2 days to reliable detections on your scene
- £200 per camera, per year
- 2 weeks risk-free guarantee · pay nothing if not accurate
- UK GDPR data residency · on-prem available
/ how it works
Plug it in.
Plug it in.
The platform does the rest.
It doesn't start with a class list or a labelling project. It starts with your stream — analyses what's there, cross-references your documentation, and proposes the detections that match your operation.
step 01
Examine the stream
The AI watches the camera feed and writes a description of the operation in plain language — what's happening, who's there, what equipment is in shot, where the work happens.
input → rtsp://*
step 02
Cross-check your context
Drop in your SOPs, training videos, sample event footage, anything. The AI reads it, watches it, and reconciles what's on the page with what's on the camera.
input → text · pdf · docx · video
step 03
Propose the detections
The AI drafts what to track, what counts as an event, and what the data table will look like. You confirm — or describe what's missing in plain English.
output → schema preview
step 04
Run, deliver, alert
Detections start running. Events stream into your dashboard. Alerts go to email, SMS, or a phone call when something matters.
output → events · alerts · calls
// accuracy maintenance, drift detection, and per-camera retraining run continuously in the background — see auto-training for how that works.
/ what's different
Built for the operation,
Built for the operation,
not the lab.
Three structural choices set Darkfield apart from every detection product before it. Together they collapse what used to be a six-month CV deployment into a forty-eight-hour conversation.
001 / authoring
The pipeline writes itself.
conventional
A computer-vision engineer designs the model, the labels, the schema. Each new task is a new sprint.
darkfield
THINK reads your documentation, inspects the stream, and authors a runnable pipeline. Onboarding is a paragraph and a camera.
002 / vocabulary
Open language, type-aware.
conventional
Closed class list. Adding "spiral mixer" requires labelled data, a retrain, and a release.
darkfield
Prompt in plain language. SEE distinguishes a spiral mixer from a planetary mixer on the first frame — no labels, no retrain.
003 / maintenance
The model is its own QA.
conventional
Accuracy drifts as the scene changes. You schedule a redeploy on a calendar, not on a signal.
darkfield
THINK audits SEE on a rolling sample. When precision drops below tolerance, it triggers a per-camera finetune — automatically.
/ first 48 hours
A model that gets better
A model that gets better
while you sleep.
From onboarding to steady-state operation, with no human in the loop after hour zero. Indicative numbers across recent partner sites.
hour 00
hour 06
hour 24
hour 48
First evaluation
THINK pulls a rolling sample of recent events and scores them against ground truth it authored. Drift detected at the edges of the work envelope, in low-light frames.
recall
0.81 → 0.83
First retrain
A per-camera finetune is launched automatically. The adapter trains on 2,800 model-curated samples in 90 minutes and hot-swaps at validation.
recall
0.83 → 0.91
Steady state
Precision and recall both inside the tolerance declared at onboarding. The eval-and-retrain loop runs continuously from this point on.
precision
0.97
// a 6–11pp recall gain in the first 48 hours · methodology, post-beta
/ ready to see it
We ship the kit.