This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
AI educational tools crave your screen space, but FieldClass AI requires less of it.
The equation for calculating the height of the tree will remain abstract until you stand in front of it. FieldClass brings the concept that you are learning about and transforms it into an experiment in nature. You read the task within a minute, put away your phone, and go and measure something real. After that, you enter your results, and your own solution to it. FieldClass checks your result, offers you a hint, and only then – shows you the correct solution.
Being a student myself, I know how hard it is to relate concepts taught in textbooks to real life. We often memorize equations without realizing where they are applicable in real life. That is why I created FieldClass AI. I aimed to make learning more practical for students, encouraging them to step out of their comfort zones and do things rather than memorize them.
It all starts with four experiments:
Trigonometry: Estimate the height of a tree, pole, or building with a cheap clinometer constructed out of a straw, protractor, and string.
Statistics: Count and classify objects within a 1 m² square and then determine their frequencies.
Plant Morphology: Measure and compare two fallen leaves.
Average Velocity: Time your friend walking on an unoccupied measured path and calculate his average velocity.
The app was designed for high school and first-year college students who have trouble seeing a STEM topic in a concrete way, and for the educator who wants a quick field exercise.
How it gets people off the screen. The mission screen has one instruction in a blue box: Read this. Then put your phone away. The read-aloud button helps you hear rather than stare. The screen pops up again once you come back with the measurements. The step tracker in the header tells you where you are (Choose, Mission, Report, Feedback), the "Bring" screen is a tick-box list for packing your bag, and the feedback screen has a closeness meter and a graph of your answer versus the answer you would get from your own measurements.
Demo
I took it outside. I estimated the height of a building from 12 m away at 45° with my eyes 1.5 m above the ground, and got 13.5 m. I also timed 20 m in 5 s (4 m/s) and measured the heights of five students (mean 160 cm).
Code
mukul9116
/
fieldclass-ai
this app enforces students to do practical learning, take whatever topic you want to study and it will guide you to learn it by doing practical
FieldClass AI
Your textbook becomes a field experiment.
FieldClass AI turns a topic you are studying into a short outdoor experiment, then helps you make sense of what you found when you come back. It runs on a small open-weight model on your own machine, so nothing you type leaves your computer.
Most AI study tools add screen time. FieldClass does the opposite: it hands you a mission you can read in under a minute, gets out of the way while you go outside and measure something real, and then checks your work when you return.
Built for the Hacktoberfest 2026 Open-Source AI Challenge, Week 1, theme Touch Grass.
What it does
- Choose a topic, a difficulty and how many minutes you have.
- FieldClass writes a short field mission: what to bring, what to do, what to write down, and a safety note.
- The mission can be read aloud…
How I Built It
The model in question is Gemma 3 1B, open weight one, working on a 8GB local machine using Ollama without any GPU. The environment around: Python, Pydantic, Streamlit, and 34 pytest tests.
At first, my prompt told the model to come up with an entire mission on its own. That resulted in the fluent but wrong solution: "the angle of the sun's rays" instead of "tan of the angle"; it excluded eye height from the formula; it cut the warning regarding the safety from the middle of the sentence; moreover, it complimented the students on what they hadn't done yet.
Therefore, I decided to change the format and not the prompt because the model knows nothing about facts.
Each experiment card includes material and procedure description, safety rules and the formula. The model calculates the correct answer based on the measurements provided by the student, gives verdict, and identifies typical mistakes: if the answer was 10.0 but it should have been 11.5, the hint reads "you might have forgotten to add your eye height". The model only generates title, objective, hint and sentence for the next challenge.
The model is never fed the right answer, so it can never reveal it, and no input of a student is ever sent to the model.
Reducing the workload of the model resulted in speeding it up. Now, a mission takes 10-17 seconds at 16 tokens per second, while previously, it took up to 50 seconds; and feedback takes 5-10 seconds. This is my personal assessment, not benchmarking.
Things that I could not do: the model often connects words ("degreesto"), it ignores any notes of the student, and I did not have a machine to test the bigger Gemma. The model is a single-profiled model in a single configuration file.
Why Does Open Innovation Matter?
Privacy and No Internet. All missions, grading, and feedback happen using the student’s personal device. The student's results and answers never reach any other server, and the application will work offline after the initial model is downloaded. It also does not cost anything for every usage.
Seeing how it is done. The local small model will be easy to see. I would test it hundreds of times, see the exact way of failure, and design around it. The closed bigger model would provide more polished sentences, and I would not notice the errors in the facts.
The tradeoff of honesty. Closed model would give better sentences. I sacrificed quality for privacy, free cost, and control and used Python for that purpose.
Prize Categories
Best Use of Gemma. Gemma 3 1B runs locally through Ollama, and the whole design, where Python owns the facts and checks what the model writes, came from testing how a small Gemma behaves.
Top comments (1)
tr.ee/dev-to