FieldClass AI
An open-source AI learning tool that converts abstract STEM textbook concepts into hands-on experiments conducted outdoors. It runs on a local small model (Gemma 3 1B via Ollama) and encourages students to leave their screens and take real-world measurements.
Timeline (last 90 days)
FieldClass AI is an open-source AI educational tool built by developer mukul9116 for the Hacktoberfest open-source AI challenge under the 'Touch Grass' theme, encouraging students to go outdoors and measure real objects to learn STEM concepts.
FieldClass AI currently has four built-in introductory experiments covering STEM knowledge from high school to freshman level: trigonometry, statistics, plant morphology, and average speed.
The workflow used by FieldClass AI is: after students read the task instructions they put away their phones to measure real objects, then return to enter their measurements and answers, the app first checks the results and provides hints, and only then reveals the correct answer.
The model used by FieldClass AI is the open-weight Gemma 3 1B, running via Ollama on a local machine with 8 GB of RAM and no GPU.
FieldClass AI's tech stack is Python, Pydantic, and Streamlit, with 34 pytest tests.
When the model was initially tasked with generating entire experiment instructions on its own, a 'fluent but incorrect' problem emerged, such as writing 'the angle of sunlight rays' instead of 'the tangent of the angle,' omitting eye height from formulas, and truncating safety instructions.
FieldClass AI's solution was to change the format rather than the prompt: each experiment card pre-contains materials, steps, safety rules, and formulas, and the model only calculates answers based on student measurement data, renders a judgment, and identifies common errors.
In FieldClass AI's design, the model never has access to the correct answer, and content entered by students is never sent to the model, thereby providing privacy protection.
After offloading the workload from the model, FieldClass takes 10 to 17 seconds per task generation (approximately 16 tokens per second), down from a previous maximum of 50 seconds, with feedback taking 5 to 10 seconds (the developer notes this is personal perception rather than a formal benchmark).
FieldClass AI has limitations: the model frequently concatenates words (e.g., 'degreesto'), tends to ignore student notes, the developer has not machine-tested larger Gemma versions, and currently only a single model profile under a single configuration file is supported.
1 more timeline events
All Facts (11)
FieldClass AI is an open-source AI educational tool built by developer mukul9116 for the Hacktoberfest open-source AI challenge under the 'Touch Grass' theme, encouraging students to go outdoors and measure real objects to learn STEM concepts.
50%UnverifiedFieldClass AI currently has four built-in introductory experiments covering STEM knowledge from high school to freshman level: trigonometry, statistics, plant morphology, and average speed.
50%UnverifiedThe workflow used by FieldClass AI is: after students read the task instructions they put away their phones to measure real objects, then return to enter their measurements and answers, the app first checks the results and provides hints, and only then reveals the correct answer.
50%UnverifiedThe model used by FieldClass AI is the open-weight Gemma 3 1B, running via Ollama on a local machine with 8 GB of RAM and no GPU.
50%UnverifiedFieldClass AI's tech stack is Python, Pydantic, and Streamlit, with 34 pytest tests.
50%UnverifiedWhen the model was initially tasked with generating entire experiment instructions on its own, a 'fluent but incorrect' problem emerged, such as writing 'the angle of sunlight rays' instead of 'the tangent of the angle,' omitting eye height from formulas, and truncating safety instructions.
50%UnverifiedFieldClass AI's solution was to change the format rather than the prompt: each experiment card pre-contains materials, steps, safety rules, and formulas, and the model only calculates answers based on student measurement data, renders a judgment, and identifies common errors.
50%UnverifiedIn FieldClass AI's design, the model never has access to the correct answer, and content entered by students is never sent to the model, thereby providing privacy protection.
50%UnverifiedAfter offloading the workload from the model, FieldClass takes 10 to 17 seconds per task generation (approximately 16 tokens per second), down from a previous maximum of 50 seconds, with feedback taking 5 to 10 seconds (the developer notes this is personal perception rather than a formal benchmark).
50%UnverifiedFieldClass AI has limitations: the model frequently concatenates words (e.g., 'degreesto'), tends to ignore student notes, the developer has not machine-tested larger Gemma versions, and currently only a single model profile under a single configuration file is supported.
50%UnverifiedFieldClass AI chooses to sacrifice the text quality of closed-source models in exchange for privacy, zero cost, and controllability, and uses Python to backstop factual calculations—a pragmatic engineering philosophy: when the model is unreliable, let deterministic code handle factual computation.
50%