Edge AI for Microcontrollers

AI that runs on
every chip.

Upload sensor data. Train an ultra-compact neural network. Download a compiled binary for your microcontroller. 95%+ accuracy on real-world benchmarks, 6x less memory than traditional ML. No cloud. No dependencies. Just pure C.

$ upload data.csv --train --target esp32 --deploy
Start Building → Log in
How it works
Four steps to on-device AI.
From raw sensor data to a hardware-locked binary in minutes. No ML expertise required.
01

Upload CSV

Drop your sensor data with labels. That's the only input we need.

02

Train

Edge V3 trains a neural network in seconds. 95%+ accuracy on real-world data. 6x less memory than traditional ML.

03

Register UIDs

Bind the model to your chip hardware IDs. It only runs on your devices.

04

Download

Get a compiled .a + .h ready to flash. Pure C, zero dependencies.

luviner-cli
$ luviner train --data sensor_data.csv --target esp32
[info] Loading dataset... 1,024 samples
[info] Detected 6 features, 3 classes
[train] [████████████████████████████] 100%
[result] Accuracy: 95.2%
[result] Model size: 1.8 KB
[result] Inference: 0.3 ms
✓ Binary compiled → model_esp32.a
Upload CSV, Train Edge V3, Compile Binary, Deploy to MCU
Edge V3
Edge V3Performance
Numbers that matter.
30x
Less energy than traditional neural networks
vs standard DNNs
95%
On UCI HAR standard benchmark (real sensor data)
UCI HAR official split
€2
Runs on the cheapest microcontrollers
total hardware cost
6x
Less memory than traditional ML models
avg ~14 KB vs 82 KB (MLP float64)
<2KB
Typical model footprint
fits any MCU
4
From sensor data to firmware
no ML expertise needed

Validated on UCI HAR public dataset — 95.0% accuracy on the standard benchmark (official split, 561 features, 2,947 test samples). Competitive with CNN and LSTM, deployable on any MCU.

See all benchmarks →
Capabilities
Beyond classification.
Built-in tools for the hardest real-world deployment challenges — from anomaly detection to automatic hardware optimization.

Anomaly Detection

Your sensor learns what "normal" looks like and detects faults, drifts, and anomalies — without any labeled fault data. Train on normal operation only.

No labeled faults needed 100% detection rate Streaming

AutoML for MCU

Specify your chip and memory constraints. Luviner automatically finds the neural network architecture that maximizes accuracy within your hardware budget.

12+ MCU profiles Custom constraints Automatic

On-Device Adaptation

Deployed models improve with ~50 new samples, directly on the chip. No retraining from scratch. No cloud connection needed.

+26% improvement 0.13 KB extra RAM On-device

Model Compression

Need to fit on a tiny chip? Luviner transfers knowledge from a large model to a compact one — achieving accuracy impossible with direct training alone.

+34% accuracy Large to tiny Same hardware

Drift Detection

Your device monitors incoming data and signals when it no longer matches the training distribution. Triggers retraining alerts automatically — no cloud needed.

No forward pass 24 bytes RAM Exclusive
Live Demo
See it in action.

Real-time motor fault detection running on a simulated ESP32. Click Play to watch.

Predictive Maintenance — ESP32
100% accuracy — 97 KB flash
Open Full Demo →
Compatibility
Every architecture. One platform.
ARM Cortex-M0 ARM Cortex-M3 ARM Cortex-M4 ARM Cortex-M7 ARM Cortex-M33 ESP32 ESP32-S3 RISC-V
Use cases
Built for the real world.
On-device intelligence for industries where latency, privacy, and power matter.

Predictive Maintenance

Detect machine failures before they happen. Vibration, temperature, current sensors — all processed on-chip.

Wearables

Gesture recognition, activity tracking, heart rate classification. On-device, no cloud dependency.

Medical Devices

ECG arrhythmia detection, SpO2 monitoring, real-time diagnostics directly on the chip.
Mesh Intelligence
MeshNew
Mesh Intelligence: your sensors collaborate.

Each sensor has its own brain. They share neural states over a lightweight mesh protocol — no cloud, no central server. The network tolerates faulty nodes, self-heals, and improves on the field without retraining.

+20%
accuracy boost vs solo nodes
24B
per message
0
cloud required
✓
optional cloud gateway

The mesh works fully offline. Optionally, one node acts as a gateway and forwards alerts to your dashboard via WiFi or LoRa — only results, never raw data.

✓ Tamper-resistant: compromised nodes are excluded automatically
✓ Self-healing: the network reconfigures when a node fails
✓ Multi-hop: information reaches the entire network
✓ Smart sharing: each node learns what information matters
✓ On-field learning: the swarm improves without retraining
✓ MCU-ready: all features compile to pure C firmware automatically
6 ENTERPRISE FEATURES
FAQ
Frequently Asked Questions
Yes. Your sensor data is used exclusively to train your model and is never shared with third parties or used to train our own models. All connections are encrypted with TLS, and you can delete your data at any time. See our Privacy Policy for details.
Absolutely. The compiled binary runs 100% on-device with zero cloud dependency. Once flashed on your microcontroller, it needs no internet connection, no API calls, and no external libraries. Pure C, completely self-contained.
On benchmark datasets we achieve 98.2% average accuracy (Iris, Wine, Cancer, Digits). On the UCI HAR standard benchmark we achieve 95%. Real-world results depend on your data quality and the complexity of your classification task. You can train and evaluate for free on the Explorer plan before committing.
Advanced hardware-level copy protection ensures your model runs only on registered devices. Each compiled binary is uniquely bound to your authorized hardware. Unauthorized devices cannot execute the model.
Binaries you've already compiled and deployed continue to work indefinitely — they're standalone and don't phone home. You just lose the ability to train new models and compile new binaries. Your account reverts to the free Evaluation plan.
No. Upload a CSV with your sensor readings and labels, click Train, and Luviner handles everything — feature extraction, network architecture, training, quantization, and C code generation. The entire process takes minutes.
Yes. AutoML for MCU lets you specify your target hardware (e.g. Cortex-M0 with 32 KB Flash and 8 KB RAM) and Luviner automatically finds the neural network architecture that maximizes accuracy while fitting within your memory budget. 12+ MCU profiles are pre-configured, or you can set custom Flash/RAM limits.
From the Blog
Technical deep dives.
Mar 14, 2026

Enterprise-Grade Mesh: 5 Features That Make Distributed AI Production-Ready

Tamper resistance, self-healing, multi-hop reach, intelligent sharing, and on-field learning — all running on 2 EUR micr...

 Mar 14, 2026

Mesh Intelligence: When Your Sensors Form a Distributed Nervous System

Each sensor has its own brain. They share neural states over a 24-byte mesh protocol. Together, they classify what no si...

 Mar 14, 2026

Anomaly Detection Without Fault Data: How Luviner Enables Predictive Maintenance from Day One

Most predictive maintenance systems need examples of every failure mode. Luviner only needs your normal operation data —...
Read all articles →
Stay in the loop
Don't miss a thing.

New benchmarks, features, and technical deep dives — straight to your inbox. No spam.

We respect your privacy. Unsubscribe anytime.

Ready to deploy AI at the edge?

Start for free. No credit card required.

Start Building →