nirvaannn

What did I do today?

• Added a todo list tool for the LLM so keep track of tasks
• Updated it to support API keys (for HackClub AI and other providers)
• Fixed tool call issues
• Tested with cloud and local models

What did I do just now?

• Added support for argparse instead of just hard-coding model names and base URLs
• Added a new subagent/bash writer tool to reduce big model context use
• Try to make it an agentic loop but ran into some issues to solve next time

What have I done since last devlog?

• Made it use the MPU6050 gyroscope to go straight
• Made it go only 95% of the cell if there was a wall in the front but then removed this due to error building up
• Tweaked the values for optimal balance between PID and sensing points
• Lots of testing in the maze with a variety of configurations (parking garage/blind driving, spirals, zigzag etc.)

Tomorrow is the real UKMARS national competition!

What did I do today?

• Used a spreadsheet to measure how much to go in the cell
• Used #define to make it easier to store these values in the code
• Changed the values and tested in the maze
• (with AI’s help) found out that it was going short due to braking compensation being applied every time compared to just at the end
• Added a bool to the calculateCellPulses() function specifying whether to apply braking compensation or not (learned about default arguments in functions)
• Found that the values from 1 cell to another in my spreadsheet don’t add to 100%
• Recalculated the values and tested (it worked!)

What have I done this time (a lot…):

• Tuned the drive distance in the flood fill loop
• Added back the 3rd sensor value at 20% of the cell from the sensing point
• Tested the robot in the maze
• 3D printed a back support to stop the robot tilting when starting drive
• Tested on a variety of cell distances and compared encoder counts in a spreadsheet
• Found that it overshot by a constant braking distance of around 60 counts each time
• More testing in the maze with updated cell calculation function incorporating braking distance
• Measured distance travelled with expected distance in a spreadsheet & tweaked the value

What did I do in this session:

• Revert a change from last time which caused malfunction
• Tuned Kp and Kd values
• Add a cell driving test
• Measured how much to drive in the cell

What did I do today:

• Replaced the 2 small models with 1 big model
• Simplified the shell tool user print so it only shows the base command

What have I been up to?

• Made it PID drive on both left and right alls
• It now only uses PID drive after 2 confirmed wall readings
• Calibrated the Kp and Kd values for the shorter distance

What did I just do?

• Finished the website with details and a footer
• Added buttons for RSVP and Slack
• Compiled it as a static HTML
• Deployed it to Netlify

Work so far:

• Setting up the Svelte project
• Creating basic containers & styles

What did I work on just now?

• Fixed a 90º bend in the LED
• Got a quote from JLCPCB
• Fixed a via issue

What did i just do?

• Restructure the PID drive flag to be an int instead of bool
• Added PID for the right wall

So now it drives smoothly as long as 1 wall is present

What did I do today?

• Routed the VCC of the LEDs
• Moved the USB port for easier access
• Added rounded corners!

I made it use the wall sensor to improve driving in a straight line using PID controller!

What did I do during this session?

• Wrote a cell-based distance driving function
• Redid it because the first one had issues
• Switched from Pico 2 to normal Pico to improve reliability

I made the app, starting with Python!

• Tested the OpenAI SDK with llama.cpp
• Added a loop
• Created tools (web search with DDGS, terminal commands)
• Added token tracking
• Used a smaller sub-model to aid in search summarization

What did I learn along the way?

• Subprocess and Python-based process management
• Streaming LLM responses

What did I do today?

• Got the mouse to bump into wall when turning 180º - this helps it calibrate itself and stay straight!
• Got the mouse to be able to return to the goal

Next I need to add a fast mode!

What did I do today?
i worked more on the maze solving loop, trying to fix some issues with wall detection

turns out, it has to go to the sensing point before reading the side walls which interferes with the sense-think-act loop!

so now it reads walls of the next cell while moving from the previous to next cell!

I did a lot today:

• Implemented positioning for the robot
• Thought about the maze loop algorithm
• Created the maze loop algorithm

I also vibe-coded an app to watch the floodfill loop from the mouse (but that’s not included in this time bc it’s AI)

What did I do today?

I wrote the floodfill algorithm in Arduino after finishing the learning slides!
I added multiple functions & it compiles!
I learned about how to structure code efficiently & also the std:queue system

Next steps: implement this with the movements of the robot to make it solve a maze!

What did I do just now?

• Learned about floodfill through UCLA IEEE document
• Made a presentation to learn about it
• Wrote the base algorithm in pseudocode, next i have to do wall checking!

I got the MPU to work and added it to the turning
Now it uses encoder-based turning with MPU for checking/compensation

What did I do today?

• Used an I2C scan to make sure my MPU shows up
• Tested the MPU with help from Claude
• integrated this into my main code
• got it to log through the Serial
• tried to use it for turning but don’t know where to put it (interrupt, loop(), core 2 etc.)

i tried to add the LEDs’ 5v pins, but need to learn how to make the wires thicker using netclasses

I added a GitHub action to compile it into a pyinstaller binary (with help from Gemini)

I tried to add the MPU 6050 gyroscope in, but it just hangs & doesn’t give any readings so i don’t know why!

I experimented with the Kakao KANANA and LG EXAONE LLMs, as they’ve worked well in the past.

They are VERY fast even at high quantizations, but lose some world knowledge compared to Gemma3n or Phi4-mini.

What have i done since then?

• Updated the KittenTTS to use the new version of the model
• Tried the Phi-4 LLM which is quite good for world knowledge!

I added SK6812 LEDs to my PCB and routed their ground and data pins

Next will be to route the 5v IN pins for them!

What did I do today?

• Tried to use the Alpha Vantage API only to find out it was very rate limited
• Got AI to help me parse the other API data
• Test it and it works!

I tried to add text to show the name of the stocks and the percentage, but something is wrong! The stocks which are down are showing as 6% up.

I’m working with AI on how to parse the CSV better to avoid this.

What did I do today:

I added an HTML version so it doesn’t need Python Turtle anymore! This means I can deploy it to the web, and also view it on systems without Python.

My next step (using HTML) would be to show the name/details of the stocks when clicked on

What did I do:

• Started by parsing the stocks from yfinance
• Calculated up/down versus opening price
• Added Python turtle to visualize
• Changed the API to avoid rate limiting
• Add it to GitHub and write a README
• Try and package it using PyInstaller

A quick session where I:

• Hardcoded the maze to see how well it drives without PID
• Added a config menu so I don’t need to keep recompiling to change mode!

I worked for 1 hour to:

• get wall following to work
• rewrite the stateful wall following function
• calibrate different turning compensation and Kp/Kd values

and now it works (kind of…)

I added Bluetooth-based configuration to my robot - no more recompiling just to change Kp or Kd values!

Experimented with different models in llama.cpp and found Gemma 3n e2B to be the best balance of performance to world knowledge