Cole Design and Development logo

Embedded AI Development

We turn sensor data into production-ready intelligence inside medical and industrial devices. We embed compact, power-efficient neural networks directly into resource-limited microcontrollers. Through quantization, pruning, and hardware-aware tuning, we pair ultra-low-power firmware with optimized models so that your devices can perform on-device Machine Learning. Typical applications include predictive maintenance, anomaly detection, and machine vision.

We build on our own open-source neural network library, written in pure C with nothing beyond the standard library, which can keep training on a device after it ships. Eight worked examples come with it, from wake word detection to predictive maintenance.

View the library on GitHub

Worked examples

Eight complete applications, each built for a microcontroller

Each example builds, trains and tests a network with the library, generates or downloads its own data, and is written up as a short technical paper with measured results. Read in order, each one asks a little more of the hardware, or of the real world, than the last.

  1. Character Recognition

    CNN

    Reads a handwritten digit from a 28×28 image. A convolutional network and a plain fully-connected one train on the same data, so what convolution buys is measured, not assumed.

    98.24%

    accuracy on 7,000 held-out digits

    100K parameters

    View example

  2. Gesture Recognition

    RNN

    Tells still, shake, tilt left and tilt right apart from a short burst of accelerometer readings: the model behind flick-to-wake on wearables and remotes.

    200 of 200

    test windows classified correctly

    388 parameters

    View example

  3. Fall Detection

    RNN · GRU · LSTM

    Watches a continuous accelerometer stream and flags a fall sample by sample, as a fall-detection pendant must, with three recurrent layer types compared on identical data.

    Every fall

    caught with no false alarms, across 20 runs of each

    337 to 1,297 parameters

    View example

  4. Wake Word Detection

    GRU

    The always-on listener in front of a voice-controlled device. Decides whether a second of real speech was the wake word, using audio feature extraction that runs in firmware.

    85%

    median accuracy on held-out speech clips

    5,121 parameters

    View example

  5. Image Classification

    CNN

    Tells a cat from a dog in an ordinary photo: the decision a camera-equipped pet door makes on its own microcontroller, trained on about 23,000 real photos.

    86%

    on a balanced test set, in 36 KB of flash

    25K parameters

    View example

  6. Object Detection

    Fully convolutional

    Finds and counts the red cells, white cells and platelets in a microscope photo of a blood smear, the job of a handheld blood analyzer.

    Every white cell

    found in the labeled test set

    16K parameters

    View example

  7. Predictive Maintenance

    FC + FFT front end

    Names which of seven faults is developing in a motor and gearbox from its vibration, while a standard vibration alarm still reads normal. Cross-compiles for an STM32U585.

    68 to 81 days

    median warning on 140 simulated machines, no false alarms

    4,792 parameters

    View example

  8. Prosthetic Hand

    FC + EMG front end

    Turns forearm muscle signals into commands for a powered hand and wrist, then fine-tunes itself to each new wearer on the device from under a minute of calibration.

    Over half

    of wrong hand movements cut by confidence-based rejection

    4,423 parameters

    View example
Read the full write-ups on GitHub

Ready to build your next intelligent device?

Let's discuss your project and how Cole Design & Development can bring your idea to life.

Talk to Us TodayExplore Our Services

Founded in 2008 · Hudson, Ohio · info@coledd.com