What is edge AI, and when does it make sense?

Edge AI runs the model on the device instead of in the cloud. It makes sense when latency, bandwidth, privacy or connectivity matter — a camera that must react in 20 ms, a sensor that has no network, or data that must never leave the site.

How small can a model get?

With quantisation, pruning and the right architecture we routinely fit keyword spotting, anomaly detection and simple vision models into a few hundred kilobytes on a Cortex-M microcontroller; richer vision runs on Jetson, Coral or an NPU-equipped SoC.

Do you also build the training pipeline?

Yes. Data collection from the devices, labelling, training, evaluation and the export/quantisation step are one pipeline, so a better model can be pushed over the air as a normal update.

How does this connect to your AI Labs work?

Directly. Edge AI is where our AI Labs research (vision, voice, local-first models) meets our embedded team — the on-device voice work behind our Sofi assistant is the same engineering.

All services

Edge AI & TinyML

Your product sees, hears and decides on its own.

Instant responses, no cloud bill, and data that never leaves the device.

When a product has to react the moment something happens, a round trip to the cloud is too slow, too costly, and stops working the instant the connection drops. We put the model on the device itself — so it responds in milliseconds, keeps working offline, and the data it reads never has to leave the customer’s hands. It is the models from our AI Labs made to fit a microcontroller, a Jetson or the NPU in your SoC.

Let’s manage Open Kanban

Computer vision running on an edge device

01 / What it solves

Products that react on their own.

See and act in real time

Spot, count and track what matters in the camera feed instantly — from a tiny Cortex-M sensor to a multi-camera Jetson rig.

Respond to voice, hands-free

Wake words, spoken commands and sound alerts handled on the device, with no audio sent anywhere.

Catch problems before they cost you

Predictive maintenance, gesture and activity recognition read straight from the sensors on board.

Keeps getting better after it ships

One loop from data to training to on-device evals and over-the-air updates, so the product improves in the field.

02 / Stack

The technology we build with.

Industry-standard, boring in the good way: tools with a future, that the next engineer will know too.

Inference runtimes

TensorFlow Lite / LiteRT
Quantised models on MCUs and mobile SoCs
ONNX Runtime
Portable inference across CPUs, GPUs and NPUs
TensorRT & DeepStream
GPU-accelerated vision pipelines on Jetson
Edge Impulse & TinyML
Rapid model iteration for constrained targets

Target hardware

Cortex-M with CMSIS-NN
Keyword spotting and anomaly detection in kilobytes
NVIDIA Jetson
Multi-camera vision at the edge
Google Coral / Hailo
Dedicated accelerators for low-watt inference
Qualcomm & NXP NPUs
On-SoC acceleration in production devices

Model pipeline

PyTorch
Training, distillation and export
Quantisation & pruning
INT8 / INT4 models that keep their accuracy
Evals on-device
Accuracy, latency and power measured on the target
OTA model delivery
Models shipped and rolled back like firmware

03 / Approach

How we build things that last.

Measured on the target

Latency, memory and current draw are measured on the real device, every model revision.

Milliwatts, not watts

Duty-cycled inference and accelerators chosen for the power budget the product actually has.

Private by design

Raw data stays on the device; only what the product needs leaves it — if anything does.

One pipeline, device to cloud

Data from the fleet feeds training; a better model ships back as an ordinary update.

04 / Standards

Best practices we don’t skip.

We adhere to industry standards and proven methods on every project — they are what makes the estimate hold.

Baseline the task with the simplest model that could work
Quantisation-aware training before post-training tricks
Accuracy evals on real device captures, not the lab set
Latency and power budgets fixed as acceptance criteria
Fallback behaviour defined for low-confidence outputs
Model versioning and rollback through the OTA channel
Drift monitoring from fleet telemetry
Privacy review of every signal that leaves the device

05 / Infrastructure

Where it runs, and how it stays up.

On-device benchmarking rigs (latency, memory, power)
Training pipeline fed by fleet data
OTA model delivery with rollback
Drift and accuracy monitoring in production

Start

Want the model on the device?

Tell us what it has to see, hear or predict — and on what hardware. An engineer writes back with an honest scope and price.

Let’s manage or explore AI Labs

Your product sees, hears and decides on its own.

Instant responses, no cloud bill, and data that never leaves the device.