Skip to content
BLUEURAL
Blueural Labs

The questions
we are working on.

Labs is where we test what actually holds up on constrained hardware. Nothing here is published yet — this is an honest list of what we are investigating and why it matters to the product.

Active themes

Each of these exists because a product decision depends on the answer.

Intrusion detection on edge hardware

How much detection can you actually run on a fanless module?

Benchmarking gradient-boosted and neural classifiers on Jetson-class hardware against public IoT and flow datasets, measuring accuracy against latency and thermal headroom rather than accuracy alone.

Edge AIIoTBenchmarking

Behavioural baselining without labels

Can a model learn one network well enough to flag the unfamiliar?

Most environments have no labelled attack data. We are working on unsupervised baselining that adapts to a single network's shape and holds a low false-positive rate as that shape drifts.

Anomaly DetectionUnsupervised

Correlating network and physical events

Does joining camera and traffic signals reduce false positives?

Testing whether pairing video analytics with network anomalies produces higher-confidence findings than either source alone — particularly for restricted-area access and after-hours activity.

Video AnalyticsSensor Fusion

Summarizing incidents in plain language

Can a small local model explain an alert usefully?

Investigating compact language models running on the appliance to turn correlated events into a paragraph an IT generalist can act on, without sending any of it to a hosted API.

Language ModelsOn-device

Write-up queue

What we intend to publish, in rough order. We would rather post a method others can reproduce than a number without one.

001Not yet published

Detection latency on edge modules

Method and benchmark harness

002Not yet published

On-device federated learning for privacy

Early exploration

003Not yet published

Baselining drift over a twelve-month window

Needs longer-running deployments

Working with us

We are interested in collaborating with universities and with organizations willing to host a longer-running deployment. Real traffic over real time is the part we cannot synthesise.

Start a conversation →

Found a vulnerability?

We run a responsible disclosure process and aim to acknowledge reports within two business days. The full terms are in the Trust Center.

Read the disclosure policy →