MiniML-Engine: A Zero-Dependency Embedded Machine Learning Framework for Microcontrollers

MiniML-Engine is an industrial-grade embedded AI framework designed for resource-constrained systems. Its core idea is "Train on PC, Run on Metal"—train models on PC then export to optimized C++ code for microcontrollers like Arduino, ESP32 (with <2KB RAM). Key features: zero external dependencies (no NumPy/PyTorch/TensorFlow), pure Python standard library implementation, math-perfect 1:1 conversion to C++ for low-memory devices. Developed by Wilner Manzanares, available on GitHub (link: https://github.com/Shuuida/MiniML-Engine) since June 3, 2026.

Embedded AI faces unique hurdles. Traditional frameworks like TensorFlow Lite/PyTorch Mobile are too bloated for microcontrollers (e.g., Arduino Uno with 2KB RAM, ESP8266). Key constraints: extremely small memory (2KB-64KB SRAM), limited compute (no GPU/FPU), real-time response needs (millisecond-level), power limits (battery-powered). MiniML addresses these by redefining "lightweight".

MiniML uses a dual-engine design. The classic ML engine (ml_runtime.py) is for table data/simple signals (low-frequency sensors). Supported algorithms:

• DecisionTreeClassifier (CART, flattened to linear arrays for O(1) stack space)
• RandomForestClassifier (Bagging, trees stored in Flash, SRAM for voting)
• MiniLinearModel (SGD, fast dot-product inference)
• MiniSVM (Hinge Loss for binary classification)
• KNearestNeighbors (in-place insertion sort to avoid dynamic RAM)
• MiniScaler (preprocess functions for normalization)

The deep learning engine (tensor.py + layers.py) handles time-series/audio/vision tasks. Key layers:

• Conv1D/2D (dynamic indexing, READ_FLOAT macro)
• SeparableConv2D (MobileNet-style, operator fusion for speed)
• ResidualBlock1D (ResNet-like identity mapping)
• MaxPool1D/2D (sliding window)
• Flatten/Linear (recursive flattening up to 4D)
• Activation/loss functions (ReLU/Sigmoid/MSELoss/CrossEntropyLoss with Clip barrier to prevent overflow)

MiniML uses internal managers for modularity:

1. ModelManager: High-level API, smart dual-core switch (uses sklearn if present, else pure Python), automated pipeline (imputation, scaling, training, predict()).
2. DataManager: Ensures static compatibility with C, flattens trees to arrays, checks input dimensions, imputes missing values.
3. ModelFactory: Instantiates models from JSON strings (avoids circular dependencies).
4. PersistenceManager: Extracts math structures to JSON (no pickle), Sklearn interop, architecture analysis (CLI), memory estimator (pre-compile resource projection).

MiniML is used in:

• Industrial edge AI: Predictive maintenance on factory equipment (vibration data analysis).
• Wearables: Heart rate anomaly detection/gait recognition (privacy & real-time).
• Agriculture IoT: Soil humidity monitoring (decision tree for irrigation control).
• Smart home: Occupancy detection (replace PIR sensors).
• Drone flight control: Attitude estimation on STM32 (millisecond control loops).

Key innovations:

1. Pure Python (zero external dependencies, only standard library).
2. Math-perfect C++ export (no precision loss between Python and C++).
3. Memory-aware design (optimized for target platform constraints).
4. Progressive complexity (start with linear models, upgrade to deep networks with consistent API).
5. Industrial stability (dimension checks, missing value handling, overflow protection).

MiniML-Engine is a breakthrough in embedded ML—proving lightweight can be industrial-grade. It offers a full workflow: PC-based Python development → one-click C++ export to hardware. This lowers embedded AI barriers. As edge computing/IoT grows, MiniML will be critical (reduces cloud dependency, faster response, better privacy). Suggestion: Developers looking to implement AI on microcontrollers should try MiniML.