The Edge AI Platform

Edge Impulse empowers ML teams to run AI at peak performance on any edge hardware, with unmatched ease and speed.

Get started

The world’s most innovative companies partnered with Edge Impulse.

# Quick integration for any Python script
import edgeimpulse as ei

# 1. Set API key from Edge Impulse project
ei.API_KEY = "ei_dae2..."
# 2. Profile model performance on target device
results = ei.model.profile(model="/path/to/model",
                            device="cortex-m4f-80mhz")
# 3. View profiling results
results.summary()
# 4. Optimize model and convert to C++ library
from ei.model.output_type import Classification
ei.model.deploy(model="/path/to/model",
                 model_output_type=Classification(),
                 output_directory="/library")

Integrate with your existing workflows and development tools

Native Python integrations with the Python SDK

Extensive Python and Node.js API bindings

NVIDIA Pretrained Models

Enterprise level Admin APIs available

AICPA SOC 2 Compliant

Edge Impulse adheres to the strictest security and privacy standards. Learn more about our security standards here.

Feature engineering for sensor data

“Edge Impulse ‘Parameters’ configuration panel showing filter and analysis settings, including scale axes, input decimation ratio, wavelet decomposition level, and wavelet type (set to bior3.9), with a purple ‘Save parameters’ button. A graph overlay titled ‘Wavelet function’ displays a waveform over time, illustrating state‑of‑the‑art DSP algorithms for feature extraction and optimized on‑device inference.”

“Edge Impulse ‘Feature importance’ panel showing a ranked list of features with horizontal bar indicators: accX L6‑zcross, accX L6‑mcross, accX L6‑n75, accX L6‑mean, and accX L6‑median, with a dropdown set to ‘All data’ and a ‘Show more’ button—helping identify which features are most relevant in the dataset.”

“Edge Impulse ‘On‑device performance’ dashboard targeting Cortex‑M4F 80 MHz, showing inferencing time of 12 ms, peak RAM usage of 1.7 K, and flash usage of 18.4 K. Below is a confusion matrix with accuracy values for different classes, illustrating real‑time analysis of on‑device performance.”

Optimize for on-device performance

“Edge Impulse EON™ Tuner interface showing three model cards side by side with performance and accuracy metrics. The highlighted card, ‘mfcc‑conv2d‑fed,’ shows 99% accuracy with donut charts for performance (latency and memory usage split between DSP, neural network, and unused resources) and a confusion matrix demonstrating high classification accuracy—illustrating how to balance feature extraction and model architecture within memory and latency constraints.”

“Edge Impulse EON™ Compiler graphic showing a neural network icon in the center with two stat boxes: ‘70% – Up to 70% less RAM’ and ‘40% – Up to 40% less ROM,’ compared to industry standard frameworks—highlighting optimized C++ model conversion and reduced memory footprint.”

“Edge Impulse Performance Calibration view showing a scatterplot of False Rejection Rate (FRR) versus False Activation Rate (FAR). Blue dots represent suggested configurations, a red dot shows default configuration before calibration, and a green dot shows a saved configuration. A performance overview panel on the right lists error details with play buttons, illustrating how to test and fine‑tune on‑device model performance for real‑world deployment.”

“Edge Impulse graphic titled ‘Build for any edge device,’ showing a central neural network icon surrounded by logos of supported hardware partners, including Silicon Labs, ALIF Semiconductor, BrainChip, Synaptics, NVIDIA, Syntiant, Nordic Semiconductor, Arduino Pro, Sony, Texas Instruments, Infineon, and Renesas—highlighting the EON™ Compiler’s ability to deploy models across diverse edge devices.”

Anomaly detection

Leverage state-of-the-art anomaly detection algorithms and choose the method that best suits your data and application.

Uncover unexpected signals

Detect anomalies in all sensor modalities with models trained solely on normal data, capable of identifying suspicious signals and indicating when an anomaly is detected, directly on device.

Detect and filter anomalous inputs

Combine anomaly detection and classification models to increase the robustness of your algorithms and ensure that you are always able to detect and filter out unexpected data inputs.

Detect visual anomalies

Build optimized models to detect anomalies in highly dimensional data such as video and image data. Leverage state-of-the-art models built to run efficiently on device.

Object detection on microcontrollers

FOMO (Faster Objects More Objects) is a ground-breaking algorithm that brings real-time object detection, tracking and counting to microcontrollers for the first time. Count objects, find the location of objects in an image, and track multiple objects in real-time.