embedded-ai-deployment

Embedded AI for Engineered Systems

Deploy AI models to embedded hardware using MATLAB® and Simulink®. This skill is written specifically for MATLAB R2026a and uses APIs, functions, and workflows introduced in that release. It covers the complete lifecycle: model creation or import, verification, compression, system-level simulation, and code generation for resource-constrained targets.

Requires MATLAB R2026a or newer. Core toolboxes: Deep Learning Toolbox, Statistics and Machine Learning Toolbox, MATLAB Coder, Embedded Coder, Simulink, and Fixed-Point Designer. Workflow-specific support packages are checked during Environment Discovery. The MATLAB and Simulink Agentic Toolkits must be available.

Workflow Pattern Selection

Determine the correct workflow pattern based on model origin and deployment target.

Decision Tree

Primary discriminator for 3P models: model size + hardware class.

Q1: What is the deployment target?
 |
 +-- Cortex-M (M33, M4, M7) ---------------------> Q2
 +-- Cortex-A/R processor or DSP (C2000, etc.) ----> Q2
 +-- x86 processor or GPU (Jetson, CUDA) ----------> Q2
      |
      Q2: Where does the AI model come from?
       |
       +-- Train from scratch in MATLAB ------------> Pattern 1  (references/pattern1/workflow.md)
       +-- Pre-trained 3P model --------------------> Q3
            |
            Q3: Route by hardware class + model size
             |
             +-- Cortex-M: always Pattern 1 import
             |     (MathWorks compression, tight sim-codegen agreement)
             |
             +-- x86 / GPU: Pattern 2 if PyTorch or LiteRT
             |     Pattern 1 import if ONNX/TF (convert to Py/LiteRT recommended)
             |
             +-- Cortex-A/R or DSP:
                   +-- Small model (< 500 KB) ---------> Pattern 1 with import path
                   +-- Large model (> 1 MB):
                        +-- PyTorch / LiteRT -----------> Pattern 2
                        +-- ONNX / TensorFlow ----------> Pattern 1 import *

* Convert to PyTorch® (.pt2) or LiteRT (.tflite) to use Pattern 2 instead.

Pattern Summary

Pattern 1 vs Pattern 2 Capability Comparison

Rule of thumb: Choose Pattern 1 for small models (< 500 KB) on lean hardware (Cortex-M, DSP) where you need MathWorks compression and tight simulation-codegen agreement. Choose Pattern 2 for larger models (> 1 MB) on high-performance hardware (x86, GPU, Cortex-A) where simulation speed is a priority and compression is done externally in Python. For Cortex-A/R and DSP targets, model size is the primary discriminator. Pattern 2 supports PyTorch (.pt2) and LiteRT (.tflite) formats. Both patterns support Simulink integration.

Common Start: Prerequisites

Regardless of pattern, always begin with these two prerequisite steps before entering the pattern-specific phases (which start at Phase 1):

1. Environment Discovery: Load references/shared/environment-setup.md
2. Project Discovery (interactive): Load references/shared/project-discovery.md

Project Discovery determines the workflow pattern via the decision tree above.

Banned Legacy Functions

Global Rules

ALWAYS

• Check toolboxes via detect_matlab_toolboxes and support packages via matlabshared.supportpkg.getInstalled before any workflow step
• If a support package is missing, ask the user to download from Add-On Explorer -- never install on their behalf
• Guide the user step-by-step -- one phase at a time
• Use rng("default") before any data splitting
• Verify numerical equivalence at each transformation step
• Generate MEX for desktop validation before generating C code for target
• Use arguments blocks in all codegen-ready functions
• Use single precision for all inference inputs
• Script-based execution: For each workflow step done in MATLAB, create a .m script file and execute it with run_matlab_file. Do NOT run ad-hoc MATLAB commands without first writing the script file. If a script needs changes, edit the script file and re-run it. This gives users full visibility into what code is being executed and enables reproducibility. IMPORTANT: run_matlab_file sets the working directory to the script's folder. Always use absolute paths (via fullfile) for model files, data, and saved outputs — never rely on pwd or relative paths.
• Pause after each workflow step: After every workflow step completes, pause and explicitly ask the user for permission to proceed to the next step. The goal is to let the user read/inspect the MATLAB scripts you created, review results, and ask questions before moving on.
• Deep Network Designer: When a model is trained in MATLAB, imported, or rebuilt as a native dlnetwork, load it in Deep Network Designer (deepNetworkDesigner(net)) so the user can visually inspect the architecture. Announce this action and wait for user acknowledgment before proceeding.
• Numerical equivalency tests (import workflows): For any import from PyTorch or ONNX:
  1. Run inference on the original 3P model to collect ground-truth reference data. Do NOT use the imported MATLAB model as reference — its custom autogenerated layers may produce incorrect outputs.
  2. Run the same inputs through the rebuilt native MATLAB model and compare against ground truth
  3. After compression, report the accuracy delta vs. the uncompressed baseline (MAE, max error, % accuracy drop). Compute these from variables in the current run — never hardcode numeric values into fprintf/disp strings, because re-running the script with different inputs or a different model will then print stale numbers.
  4. Run tests to validate numerical equivalence between: compressed model in MATLAB, compressed model in Simulink, and final generated code
• Test count proposal: Before running numerical equivalency tests, propose how many tests you plan to run and explain why (considering model complexity, output range, class count, etc.). Wait for user agreement or correction before proceeding.
• Code generation report: After code generation is complete and the project is done, open the code generation report (open(reportPath) or web(reportPath)) so the user can inspect the generated code, warnings, and metrics.
• Look up function signatures from MATLAB's help or the online reference page, not from this skill. Argument lists, name-value pair (NVP) defaults, and supported-layer enumerations live in MATLAB's help <function> output and on the function's reference page. Use those as the source of truth instead of any inline parameter table in this skill — inline tables go stale across releases and burn context. This skill only flags name-value arguments that materially change the recipe (e.g., ValidationThreshold for accuracy-budgeted pruning). Lookup procedure:
  1. First try help <function> in the live MATLAB session. Fast and reflects the actually-installed release of the toolbox or support package.
  2. If the function is not found at all (which <func> returns "not found") on a system that has the relevant support package installed, the support package is likely on a stale build. Ask the user to update via Add-On Explorer rather than working around the missing function.
• Compression decision flow: At the start of Phase 5 (Pattern 1), load references/pattern1/compression-decision.md and walk the user through the question flow (hardware + Simulink availability, primary goal, retraining tolerance). Pick the compression and code generation path based on the answers. Compression is not mandatory and the optimal combination of pruning, projection, and quantization depends on the goal — for example, on Cortex-M with a latency-bound LSTM model, the float32 path with CMSIS-DSP outperforms the quantized path because CMSIS-NN provides no INT8 kernel for recurrent layers.

ASK FIRST

• Before each phase transition: "Is this step relevant to your project?"
• Before data splitting: existing train/val/test splits?
• Before model selection: problem type and constraints
• Before Simulink: existing Simulink model?
• Before quantization: hardware numeric capabilities (FP vs FXP)
• Before code generation: target deployment hardware
• Before compression and code generation (Pattern 1): walk the user through the decision flow in references/pattern1/compression-decision.md — hardware target + Simulink availability, primary goal, retraining tolerance. The answers determine the compression techniques and the code-replacement library to use.

NEVER

• Present the entire workflow at once
• Skip Environment Discovery or Project Discovery
• Open, load, or inspect user data before Project Discovery is confirmed
• Use banned legacy functions
• Assume toolbox or support package availability without checking
• Install support packages on the user's behalf
• Promise hardware-agnostic performance or "deploy anywhere"
• Generate DAGNetwork, SeriesNetwork, or network objects
• Run MATLAB commands directly in the MCP server without creating a script file first
• Skip numerical equivalency testing when importing 3P models
• Proceed to the next workflow step without explicit user permission
• Apply compression without first walking the user through the decision flow in compression-decision.md
• Use the imported model (with custom autogenerated layers) as numerical ground truth — always validate against the original 3P model
• Pass a [C × T] array with format "CBT" to a sequence model — always reshape to [C × 1 × T] for single-sequence inference
• Pass a dlnetwork to prepareNetwork — in R2026a the function takes a dlquantizer object (prepareNetwork(quantObj)) and mutates it in place. The legacy net = prepareNetwork(net) form is no longer defined

---

MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See www.mathworks.com/trademarks for a list of additional trademarks. Other product or brand names may be trademarks or registered trademarks of their respective holders.