Physical AI on edge devices
Give robots, drones, vehicles and industrial machines AI that sees, decides and acts in real time, running on the device itself rather than in the cloud.
Edge AI · Any processor
AI that runs on the device.
On every XPU.
EdgeAI XPU builds the software that takes AI models from the cloud onto edge hardware, and makes them fast, small and power-efficient on whatever processor the device has.
- NPU
- GPU
- DSP
- CPU
- MCU
Our objectives
Where we're headed
AI on every XPU
Make AI models run fast, small and power-efficient on whatever processor a device has: NPU, GPU, DSP, CPU or MCU.
What we do
From trained model to shipping device
01
Optimize
Quantization, pruning and graph-level tuning that shrink models to fit edge memory and power budgets while keeping accuracy.
02
Map to the XPU
Compile and partition each model across the NPU, GPU, DSP and CPU on the chip, so every part of the network runs where it runs best.
03
Deploy & measure
Run on real hardware, profile latency and energy, and ship a runtime that behaves the same in the field as it did in the lab.
Free tool
XPU Lens
Look inside any AI model
Open a model and explore it block by block, down to every layer: inputs and outputs, weights, settings and estimated MAC cost. It runs in your browser, so your model is never uploaded.
Open XPU LensThe models you already have
ONNX, TFLite, Keras, PyTorch and TensorFlow. No account and nothing to install, and there are sample models to try first.
XPU Live: speak, and watch it run on the edge
Ask a question out loud. A denoiser, speech recognition, a small language model and a voice answer it on our own edge AI machine, with no cloud AI, and each stage shows the processor it ran on and how long it took.
Open XPU LivePick the model for every stage
Choose from 15 open models across the four stages, including Whisper, Moonshine, Qwen, SmolLM2 and Kokoro, on the GPU or the CPU. You can also run one recording through each denoiser and hear the difference.
XPU Vision: a robot’s eyes, on the edge
Turn on your camera, or bring a picture or a video. Open models find objects, measure depth and follow body pose frame by frame, and a vision-language model answers your typed or spoken question about the scene, pointing at what it means.
Open XPU VisionQwen3-VL and Google’s open Gemma
Eight open vision models: RT-DETR v2, Grounding DINO, Depth Anything V2, ViTPose, Qwen3-VL 2B and 4B, Gemma 3 4B, and Gemma 3n, built like Gemini Nano for devices.
Why the edge
Faster, private, and always on
- Low latency. Decisions happen on the device, with no round trip to a server.
- Privacy by design. Audio, video and sensor data stay where they were captured.
- Works offline. Products keep working without a network connection.
- Lower cost. Less cloud compute and bandwidth for every unit shipped.
Leadership
Founders
SA
Surbhi Agarwal
Founder & CEO
Leads EdgeAI XPU's strategy, operations and partnerships.
AK
Ayush Kansal
Co-founder
Builds the software behind EdgeAI XPU's platform and products.
Bringing AI to your device?
Tell us about your hardware and your model. We'll get back to you.
surbhi@edgeaixpu.com