Published July 17, 2026 © GPL3+

Math Toucan Core – Train Your Brain, Anytime

A tiny offline LLM device that generates simple math challenges and solutions to keep your brain sharp anytime.

AdvancedShowcase (no instructions)Over 1 day48

Math Toucan Core – Train Your Brain, Anytime

Things used in this project

Hardware components

M5Stack Core2 v1.3

M5Stack LLM (large language model) Module Kit (AX630C)

Lego Wild Animals: Tropical Toucan

Software apps and online services

M5Stack UIFlow 2.0

Hand tools and fabrication machines

3D Printer (generic)

Story

Math Toucan Core is a pocket‑sized math trainer built on the M5Stack Core2 and the local LLM Unit. With a single button press, it generates simple math problem: Think about the problem, try solving it yourself, and when you're ready, press another button to reveal the solution.

It’s a tiny device designed to keep your brain sharp through quick, repeatable practice — no internet, no distractions, just pure mental training. Whether you have one minute or ten, Math Toucan Core helps you stay in math shape, one small challenge at a time.

Math Toucan Core

Concept & Goals

The Math Toucan Core is an unusual project that connects two worlds that, at first glance, seem to have little in common: the world of microcontrollers and the world of large language models (LLMs). This combination is exactly what makes the system exciting, because it demonstrates how flexibly LLM can be used—not just as an answer machine, but as a creative partner that invents new thinking challenges and stimulates the brain.

The Math Toucan Core has a clear goal: to make the user think.

Instead of repeating pre‑made math exercises from textbooks, the system constantly invents new problems. Some are simple, some surprising and some even a bit odd.

The user is encouraged to solve the tasks mentally first, before asking the LLM for help. This creates a natural rhythm between personal cognitive effort and machine support. The system becomes a kind of mental fitness device:

It challenges
It surprises
It forces you to rethink
It keeps the brain active

And because the tasks are not taken from a fixed catalog but generated dynamically, the training stays varied and engaging.

Example Math Problem

On the technical side, the Math Toucan Core shows how LLMs can collaborate with microcontrollers. Normally, microcontrollers are seen as rigid systems:

limited memory
limited computing power
clearly defined functions

By connecting them with an LLM, this rigid system suddenly becomes flexible. The microcontroller becomes a platform that sends LLM requests, generates tasks, processes answers, and interacts with the user.

The project demonstrates:

how LLM can make small devices smarter
how microcontrollers can produce dynamic content
how traditional embedded systems can be expanded through LLMs
how programming becomes more adaptable and creative with LLMs

The Math Toucan Core is therefore an example of how LLM pushes the boundaries of classic hardware.

User Interaction

After powering on the device, the start screen appears, and during the setup phase the LLM Unit reads aloud the welcome message, the instruction text, and the start inmstrucitions. Once this initialization is complete, the interface changes to the main menu page, where the user can begin interacting with the system.

Start Page

Menu Page

The user finds four buttons on the start page, and each button triggers the LLM with a specific prompt:

1 to 100 s: Write a very simple math problem using only addition or subtraction. Use numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.
1 to 100 w & s: Write a very simple math problem using only addition or subtraction. Use everyday objects and numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.
1 to 100 d: Write a very simple math problem using only addition or subtraction. Use numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.
1 to 100 w & d: Write a very simple math problem using only addition or subtraction. Use everyday objects and numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.

After a button is pressed, the interface switches to the problem page, where the mathematical problem generated by the LLM is displayed and read aloud. Because the model needs time to create the exercise, the generation process may take a moment before the final problem appears.

Problem Page

To display and read out the solution to a mathematical problem, you press the middle red button. Depending on the selected problem category, the Math Toucan Core generates either a simple or a detailed solution. It uses one of the twoe prompts:

Develop a simple solution for an elementary school student. Provide only the solution.
Develop a simple solution for an elementary school student. The solution should consist of one to four sentences. Provide the solution.

The right red button generates the next mathematical problem, providing a fresh challenge for the user. The left red button takes you back to the main page, where all core functions of the Math Toucan Core can be accessed.

On the main page, pressing the left red button triggers the system to read aloud the introduction text and the start instructions, giving the user a clear overview of how everything works. When the middle red button is pressed, the Math Toucan Core reads out an LLM‑generated definition of addition and subtraction—each time formulated slightly differently, depending on how the LLM interprets the prompt.

Conclusion & Problems

The Math Toucan Core demonstrates how easy it has become to integrate large language models (LLM) into modern projects and how this fundamentally changes user interaction. Instead of programming every possible use case manually, the LLM takes over much of the content logic and responds flexibly to the user’s behavior. This dynamic approach makes the system more versatile and engaging: the LLM adapts, reformulates content, generates spontaneous tasks, and expands the capabilities of a simple microcontroller far beyond traditional, fixed functions.

A crucial aspect: The LLM does not always provide correct solutions.

Sometimes the calculation is wrong, sometimes the explanation is incomplete, sometimes the task itself is a bit strange. But instead of being a flaw, this is a feature. It shows the user where the limits of LLMs lie:

mathematical precision is not guaranteed
logical errors can occur
the AI can miscalculate
some tasks may be unclear

This imperfection makes the system valuable. The user must think critically, verify, question—and that strengthens the brain even more.

Further Examples

Code

math-toucan-core.py

import os, sys, io
import M5
from M5 import *
import m5ui
import lvgl as lv
from module import LlmModule
import time



page_start = None
page_problem = None
page_menu = None
page_start_spinner = None
page_menu_buttonmatrix = None
page_problem_textarea = None
page_start_label_title = None
page_start_label_load = None
page_start_label_name = None
llm_0 = None


saytext = None
question = None
is_answer_finish = None
answer = None
operation = None
k = None

# say a welcome message during setup
def Say_Welcome_Text():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  saytext = 'Welcome to the Math Jungle. I am Dietmar, your colorful math toucan. Together, we will practice addition and subtraction at your own pace.'
  print(saytext)
  llm_0.melotts_inference(llm_0.get_latest_melotts_work_id(), saytext, 0, 'melotts_inference')
  page_start_label_load.set_flag(lv.obj.FLAG.HIDDEN, True)
  page_start_label_name.set_flag(lv.obj.FLAG.HIDDEN, False)
  time.sleep(10)

# generate a maths problem if the button labelled "1 to 100 s" is pressed.
def Generate_Category1_Problem():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  question = 'Write a very simple math problem using only addition or subtraction. Use numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.'
  print(question)
  llm_0.clear_response_msg_list()
  llm_0.llm_inference(llm_0.get_latest_llm_work_id(), question, 'llm_inference')
  page_problem_textarea.set_text(str(get_model_response()))
  Read_Textarea_Text()

# get the larg language model response
def get_model_response():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  is_answer_finish = False
  answer = ''
  while not is_answer_finish:
    llm_0.update()
    for k in (llm_0.get_response_msg_list()):
      if (k['work_id']) == (llm_0.get_latest_llm_work_id()):
        answer = (str(answer) + str(((k['data'])['delta'])))
        is_answer_finish = (k['data'])['finish']
    llm_0.clear_response_msg_list()
  return str(answer)

# say an instruction message during setup
def Say_Instructions_Text():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  saytext = 'I will pick out the right problems for you and show you just how much fun numbers can be. If something does not work out, we will just flutter on until we get it right.'
  print(saytext)
  llm_0.melotts_inference(llm_0.get_latest_melotts_work_id(), saytext, 0, 'melotts_inference')
  time.sleep(9)

# say a message to start the game during setup.
def Say_Start_Text():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  saytext = 'Listen carefully, choose a category, and our adventure will begin. But sometimes I get confused.'
  print(saytext)
  llm_0.melotts_inference(llm_0.get_latest_melotts_work_id(), saytext, 0, 'melotts_inference')
  time.sleep(6)

# generate a maths problem if the button labelled "1 to 100 w & d" is pressed.
def Generate_Category2_Problem():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  question = 'Write a very simple math problem using only addition or subtraction. Use everyday objects and numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.'
  print(question)
  llm_0.clear_response_msg_list()
  llm_0.llm_inference(llm_0.get_latest_llm_work_id(), question, 'llm_inference')
  page_problem_textarea.set_text(str(get_model_response()))
  Read_Textarea_Text()

# read aloud the problem or solution
def Read_Textarea_Text():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  if not not len(page_problem_textarea.get_text()):
    llm_0.melotts_inference(llm_0.get_latest_melotts_work_id(), page_problem_textarea.get_text(), 500, 'melotts_inference')

# generate a maths problem if the button labelled "1 to 100 d" is pressed.
def Generate_Category3_Problem():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  question = 'Write a very simple math problem using only addition or subtraction. Use numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.'
  print(question)
  llm_0.clear_response_msg_list()
  llm_0.llm_inference(llm_0.get_latest_llm_work_id(), question, 'llm_inference')
  page_problem_textarea.set_text(str(get_model_response()))
  Read_Textarea_Text()

# generate a maths problem if the button labelled "1 to 100 s" is pressed.
def Generate_Simple_Solution():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  llm_0.llm_inference(llm_0.get_latest_llm_work_id(), (str((page_problem_textarea.get_text())) + str('Develop a simple solution for an elementary school student. Provide only the solution.')), 'llm_inference')
  page_problem_textarea.set_text(str(get_model_response()))
  Read_Textarea_Text()

# generate a maths problem if the button labelled "1 to 100 w & d" is pressed.
def Generate_Category4_Problem():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  question = 'Write a very simple math problem using only addition or subtraction. Use everyday objects and numbers between 1 and 100. Make the problem short and easy to read. Only write the problem, not the answer.'
  print(question)
  llm_0.clear_response_msg_list()
  llm_0.llm_inference(llm_0.get_latest_llm_work_id(), question, 'llm_inference')
  page_problem_textarea.set_text(str(get_model_response()))
  Read_Textarea_Text()

# Describe this function...
def Generate_Detailed_Solution():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  llm_0.llm_inference(llm_0.get_latest_llm_work_id(), (str((page_problem_textarea.get_text())) + str('Develop a simple solution for an elementary school student. The solution should consist of one to four sentences. Provide the solution.')), 'llm_inference')
  page_problem_textarea.set_text(str(get_model_response()))
  Read_Textarea_Text()

# Describe this function...
def Generate_Operation_Explanation():
  global saytext, question, is_answer_finish, answer, operation, k, page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0
  saytext = 'One jungle moment, please! I’m flapping around to find the right definitions.'
  print(saytext)
  llm_0.melotts_inference(llm_0.get_latest_melotts_work_id(), saytext, 0, 'melotts_inference')
  question = 'Explain in a short and simple way what addition and subtraction is, so that an elementary school student can easily understand it. The explanation should consist of one to four sentences.'
  print(question)
  llm_0.llm_inference(llm_0.get_latest_llm_work_id(), question, 'llm_inference')
  saytext = get_model_response()
  print(saytext)
  llm_0.melotts_inference(llm_0.get_latest_melotts_work_id(), saytext, 0, 'melotts_inference')


def btnA_wasClicked_event(state):
  global page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0, saytext, question, is_answer_finish, answer, operation, k
  print('BtnA')
  Speaker.tone(2000, 50)
  if operation == 4:
    Say_Instructions_Text()
    Say_Start_Text()
  if operation < 4:
    operation = 4
    page_menu.screen_load()
  print(operation)


def btnB_wasClicked_event(state):
  global page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0, saytext, question, is_answer_finish, answer, operation, k
  print('BtnB')
  Speaker.tone(2000, 50)
  if operation==0:
    Generate_Simple_Solution()
  elif operation==1:
    Generate_Simple_Solution()
  elif operation==2:
    Generate_Detailed_Solution()
  elif operation==3:
    Generate_Detailed_Solution()
  else:
    Generate_Operation_Explanation()


# call the problem function and page
def page_menu_buttonmatrix_value_changed_event(event_struct):
  global page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0, saytext, question, is_answer_finish, answer, operation, k
  Speaker.tone(2000, 50)
  print(page_menu_buttonmatrix.get_selected_button())
  if page_menu_buttonmatrix.get_selected_button()==0:
    operation = 0
    page_problem.screen_load()
    Generate_Category1_Problem()
  elif page_menu_buttonmatrix.get_selected_button()==1:
    operation = 1
    page_problem.screen_load()
    Generate_Category2_Problem()
  elif page_menu_buttonmatrix.get_selected_button()==2:
    operation = 2
    page_problem.screen_load()
    Generate_Category3_Problem()
  elif page_menu_buttonmatrix.get_selected_button()==3:
    operation = 3
    page_problem.screen_load()
    Generate_Category4_Problem()
  else:
    operation = 4
    page_menu.screen_load()
  print(operation)


def btnC_wasClicked_event(state):
  global page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0, saytext, question, is_answer_finish, answer, operation, k
  print('BtnC')
  Speaker.tone(2000, 50)
  if operation == 0:
    operation = 0
    page_problem.screen_load()
    Generate_Category1_Problem()
  if operation == 1:
    operation = 1
    page_problem.screen_load()
    Generate_Category2_Problem()
  if operation == 2:
    operation = 2
    page_problem.screen_load()
    Generate_Category3_Problem()
  if operation == 3:
    operation = 3
    page_problem.screen_load()
    Generate_Category4_Problem()
  print(operation)


def page_menu_buttonmatrix_event_handler(event_struct):
  global page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0, saytext, question, is_answer_finish, answer, operation, k
  event = event_struct.code
  if event == lv.EVENT.VALUE_CHANGED and True:
    page_menu_buttonmatrix_value_changed_event(event_struct)
  return

def setup():
  global page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0, saytext, question, is_answer_finish, answer, operation, k

  M5.begin()
  Widgets.setRotation(1)
  m5ui.init()
  page_start = m5ui.M5Page(bg_c=0xffffff)
  page_problem = m5ui.M5Page(bg_c=0xffffff)
  page_menu = m5ui.M5Page(bg_c=0xffffff)
  page_problem_textarea = m5ui.M5TextArea(text="Text", placeholder="Placeholder...", x=2, y=2, w=316, h=236, font=lv.font_montserrat_14, bg_c=0xffffff, border_c=0xe0e0e0, text_c=0x212121, parent=page_problem)
  page_start_spinner = m5ui.M5Spinner(x=110, y=70, w=100, h=100, anim_t=10000, angle=180, bg_c=0xe7e3e7, bg_c_indicator=0x2193f3, parent=page_start)
  page_menu_buttonmatrix = m5ui.M5ButtonMatrix(["1 to 100 s","1 to 100 w & s","\n","1 to 100 d","1 to 100 w & d"], x=1, y=1, w=318, h=237, target_textarea=None, parent=page_menu)
  page_start_label_title = m5ui.M5Label("Math Toucan Core", x=48, y=20, text_c=0x000000, bg_c=0xffffff, bg_opa=0, font=lv.font_montserrat_24, parent=page_start)
  page_start_label_load = m5ui.M5Label("loading", x=10, y=200, text_c=0x000000, bg_c=0xffffff, bg_opa=0, font=lv.font_montserrat_12, parent=page_start)
  page_start_label_name = m5ui.M5Label(" I am Dietmar, your colorful math toucan. ", x=11, y=200, text_c=0x000000, bg_c=0xffffff, bg_opa=0, font=lv.font_montserrat_14, parent=page_start)

  BtnA.setCallback(type=BtnA.CB_TYPE.WAS_CLICKED, cb=btnA_wasClicked_event)
  BtnB.setCallback(type=BtnB.CB_TYPE.WAS_CLICKED, cb=btnB_wasClicked_event)
  page_menu_buttonmatrix.add_event_cb(page_menu_buttonmatrix_event_handler, lv.EVENT.ALL, None)
  BtnC.setCallback(type=BtnC.CB_TYPE.WAS_CLICKED, cb=btnC_wasClicked_event)

  page_start_label_name.set_flag(lv.obj.FLAG.HIDDEN, True)
  page_start_label_load.set_text(str((str('loading ') + str('openbuddy-llama3.2-1B-ax630c'))))
  page_start.screen_load()
  llm_0 = LlmModule(2, tx=14, rx=13)
  while not (llm_0.check_connection()):
    time.sleep(1)
  llm_0.sys_reset(True)
  llm_0.audio_setup(cap_volume=0.5, play_volume=0.15, request_id='audio_setup')
  # llama3.2-1B-p256-ax630c
  llm_0.llm_setup(prompt='You are a math teacher for elementary school students.', model='openbuddy-llama3.2-1B-ax630c', enoutput=True, enkws='', max_token_len=254, request_id='llm_setup')
  llm_0.melotts_setup(model='melotts-en-default', input='tts.utf-8.stream', enoutput=False, enkws='', request_id='melotts_setup')
  time.sleep_ms(500)
  Say_Welcome_Text()
  Say_Instructions_Text()
  Say_Start_Text()
  operation = 4
  page_menu.screen_load()


def loop():
  global page_start, page_problem, page_menu, page_start_spinner, page_menu_buttonmatrix, page_problem_textarea, page_start_label_title, page_start_label_load, page_start_label_name, llm_0, saytext, question, is_answer_finish, answer, operation, k
  M5.update()


if __name__ == '__main__':
  try:
    setup()
    while True:
      loop()
  except (Exception, KeyboardInterrupt) as e:
    try:
      m5ui.deinit()
      from utility import print_error_msg
      print_error_msg(e)
    except ImportError:
      print("please update to latest firmware")

Credits

esferatec

2 projects • 0 followers

Math Toucan Core – Train Your Brain, Anytime

Things used in this project

Hardware components

Software apps and online services

Hand tools and fabrication machines

Story

Concept & Goals

User Interaction

Conclusion & Problems

Further Examples

Schematics

Setup

Say Functions

Model Response

Read Text

Problem Functions

Solution Functions

Explanation Function

Code

math-toucan-core.py

Credits

esferatec

Comments

Embed the widget on your own site

Math Toucan Core – Train Your Brain, Anytime

Math Toucan Core – Train Your Brain, Anytime

Things used in this project

Hardware components

Software apps and online services

Hand tools and fabrication machines

Story

Concept & Goals

User Interaction

Conclusion & Problems

Further Examples

Schematics

Setup

Say Functions

Model Response

Read Text

Problem Functions

Solution Functions

Explanation Function

Code

math-toucan-core.py

Credits

esferatec

Comments

Related channels and tags