// Continuous AI Operations Platform
The operating system for vision operations.
LexData takes a vision model through its whole life: verified labels, training on your footage, alerts from live feeds, and retraining when the world changes underneath it.
SOLARpanel_tablepanel+41,33720,882ready
Wind Turbine Inspectionturbineblade+32,03317,483ready
Real Thermal (Main Project)PersonVehicle+35,29434,761ready
Scaffolding Monitoringscaffold_standardledger+101,106107,162ready
Retail Stores - Store camera viewrack_sectionperson+555813,889ready
Powerline-transformer inspectioninsulatorrust_corrosion+47712,904ready
Pipeline_damage_detectionrustDent+32035,802ready
Drone Inspection of pipelineRustHuman+21581,641ready
tractor navigationTree RowDrivable Path+42312,042readySOLAR1,337 frames · 20,882 labelsreadyWind Turbine Inspection2,033 frames · 17,483 labelsreadyReal Thermal (Main Project)5,294 frames · 34,761 labelsreadyScaffolding Monitoring1,106 frames · 107,162 labelsreadyRetail Stores - Store camera view558 frames · 13,889 labelsreadyPowerline-transformer inspection771 frames · 2,904 labelsreadyPipeline_damage_detection203 frames · 5,802 labelsreadyDrone Inspection of pipeline158 frames · 1,641 labelsreadytractor navigation231 frames · 2,042 labelsready01 Annotate
Annotate
Type what to look for, in plain words. Lexi puts a box on every frame, a person checks each one, and the labels come back on your own footage at up to 99.9% accuracy, ready to train.
- Boxes, polygons, points and tagsFour annotation modes on every frame, with instance tracking across frames when a project needs it.
- Checked by a personA QA pass on every label before it trains anything.
- Import COCO, YOLO or CVATLabels from a previous vendor come in as they are.
- Footage from S3, Drive or uploadPoint at the bucket; nothing is re-encoded.
- Export the dataset as COCOYour labels are yours to take.
- Tracking across framesOne object, one id, every frame it appears in.
02 Alert
Alert
Type a rule in plain words. Every camera is watched for it, and the frame reaches the right person while there is still time to act.
- Rules written in a sentenceNo query language, no threshold sliders.
- Severity and cooldownsEach rule carries its severity and how long it waits before firing again.
- Slack, email or a webhookThe frame goes where your team already looks.
- Approved before it goes liveA rule is a card you read and switch on.
- Fires on your own siteWith the edge runner, the alert never waits on the cloud.
03 Insight
Insight
Ask any question about your footage. Every detection is kept and tied to the asset, the place and the date, so what comes back is an answer you can act on, with the frames behind it.
- Ask in plain wordsA question, not a dashboard.
- Every detection keptNothing is sampled away.
- Evidence attachedEvery answer points at the frames it came from.
- Filter by asset, place or dateAsk about one site, one week, one camera.
04 Operate
Operate
The model is watched after launch. Frames it is unsure of come back for a person to check, the corrections retrain it, and the new version rolls out with no downtime.
- Uncertain frames returnThe model sends back what it doubts; nothing else leaves the site.
- Corrections retrainWhen enough corrections land, a new version is trained on them.
- Versioned modelsEvery version keeps what it was trained on.
- Rolls out with no downtimeThe cameras never go dark for a new version.
05 Deploy
Deploy
Runs on the cameras you already have. In the cloud, on your own servers, or at the edge next to the recorder, with SSO and role based access from day one.
- Any RTSP feed or recorderNo new cameras and no new hardware.
- Cloud or your serversManaged by us either way, or run the exported model yourself.
- Runs at the edgeA runner on your network watches the recorder and its cameras.
- Exports as ONNX, PyTorch or TorchScriptWhat we train, you can take.
- Eight device presetsDescribe the device, not the architecture: from a Raspberry Pi to a GPU server.
- SSO and role based accessYour identity provider, your roles.
labels on this frame
the project
this batch
frame 467
frame 717
frame 968
frame 1218
frame 1468
frame 1718
frame 1968
frame 2218
frame 2468
frame 2718
frame 29613
frame 32116
the event
today
Where do we see a vehicle?
answer
In 3,154 frames of 5,294, all of them on the perimeter camera.
9,886 detections kept, every one from a single camera run, ffe96b4f, 9 August 2026. The two walkway runs of 10 August never see one.
counted from
the footage
asked before
Personframe 1450✓ confirmed
Vehicleframe 1490needs review
Vehicleframe 1530needs review
Vehicleframe 1570needs review
Vehicleframe 1610needs review
Vehicleframe 1650needs review
Vehicleframe 1690needs review
Vehicleframe 1730needs review
Vehicleframe 1770needs review
Vehicleframe 1810needs review
Vehicleframe 1850needs reviewwhat happens next
- 01Corrections join the training set
- 02A new version trains on them
- 03It rolls out with no downtime
the next version
the model now
cameras watched
18
recorders
5
model version
v1
inference
on the LAN
last alert
20:14
state
live
who can sign in
SSO and role based access
Your identity provider, your roles.
exports
PT, ONNX, TorchScript
Eight device presets, from a Raspberry Pi to a GPU server.
where the same model runs
Cloud
The model runs in the LexData cloud.
Your servers
Managed by us, or run by you.
Edge
A runner on your network, next to the recorder.
device presets
the streams it watches
Where it runs.Cloud, your servers, or the edge.
Cloud
Nothing to run.
Where the model runs
LexData cloud
Where footage lives
Uploaded to LexData
What leaves the site
Footage you upload
How alerts fire
Slack, email, webhook
How models export
PT, ONNX, TorchScript, eight device presets
Who can sign in
SSO and role based access
Your servers
Managed by us, or run by you.
Where the model runs
Your servers, managed by us or run by you
Where footage lives
Stays on your servers
What leaves the site
Nothing you do not send
How alerts fire
Slack, email, webhook
How models export
Same
Who can sign in
Same
Edge
A runner beside the recorder.
Where the model runs
A runner on your network
Where footage lives
Stays on site
What leaves the site
Only frames the model doubts
How alerts fire
Locally, then Slack, email, webhook
How models export
Same
Who can sign in
Same
The blueprint
How it all connects.
This is LexLoop.
The circuit runs whether or not anyone is watching it
01 · Annotate
Everything you have, in the door.
Footage from buckets and drives; labels from a previous vendor, imported as they are. LexAnnotate turns them into expert-verified training data.
Type what to look for, in plain words.
02 · Deploy
The model moves in with the cameras.
The edge runner watches on your LAN, one NVR and its cameras at a time. Cloud, your own servers, or the edge: the same loop, wherever it is allowed to live.
Runs on the cameras you already have.
03 · Alert
The frame goes where the team looks.
LexAlert fires locally, straight to Slack, email, or a webhook. Each rule carries its severity and how long it waits before firing again.
Type a rule in plain words.
04 · Operate
Only uncertain frames leave.
Frames the model doubts come back for human review, become verified corrections, and retrain the model. The circuit closes on its own.
The model is watched after launch.
05 · Insight
Every wire reports to LexInsight.
Every detection is kept, tied to the asset, the place, and the date. Ask what happened, in plain language, and make the call on evidence.
Ask any question about your footage.
The software.
Three products, one loop. Most teams start with LexAnnotate and add the rest once the model is live.
LexAnnotate
Describe what matters. Lexi labels it, and a person checks every frame.
Boxes · Polygons · Points · TagsTracking · Thermal · LiDAR
LexAlert
Rules written in a sentence, approved as a card, fired from your own cameras.
Slack · Email · Webhook · Edge-local
LexInsight
Every detection kept and answerable in plain language, so decisions run on evidence.
Asset · Place · Date · Evidence attached
Trusted in the field by
FLY GUYS
HELMERICH & PAYNE
QUIDIENT
FLY GUYSHELMERICH & PAYNEQUIDIENTIndustries
The loop runs. You run the business.
Somewhere on a feed you own, the world just changed. A camera was nudged, a season turned, a spec moved. The loop catches it, routes the doubtful frames through human review, retrains on the corrections, and redeploys - before you would have noticed anything drifted. Labels in, alerts out, accuracy holding in between. That is the whole idea.
Enterprise? Contact sales