Inspiration
A student can make one wrong move in math and carry it through an entire page. We wanted to catch that first mistake while there is still time to help.
What it does
Teachify reads handwritten math from a phone photo or desk camera, identifies the first incorrect step, and gives the student a hint instead of the answer. In class, it shows teachers who needs help and what question to ask.
How we built it
flowchart LR
A[Capture work] --> B[Read each line]
B --> C[Check the math]
C --> D[Give a hint]
D --> E[Alert the teacher]
Gemini reads the writing, SymPy checks the math, and an AI layer phrases the hint and teacher brief. The app runs on Google Cloud Run. We also designed a camera tripod and fine-tuned a handwriting reader, which is not yet connected to the live app.
Challenges we ran into
Handwritten math is difficult to read reliably. We also had to avoid treating every line after one mistake as a new error, and design a classroom option that does not depend on students using phones.
Accomplishments that we're proud of
Our fine-tuned reader reached 76.5% exact match on 1,000 held-out handwritten lines, with the benchmark limits noted in our results. On 29 labeled images, line-by-line checking left all 18 valid steps after an earlier mistake unflagged.
What we learned
Checking each line in context produces fairer feedback. A narrow fine-tune can also be inexpensive, but benchmark gains need careful qualification.
What's next for Teachify
Connect the fine-tuned reader to the live app, test the camera setup in a classroom, and add the accounts, storage, and privacy review needed for a school pilot.
Built With
- claude
- coud
- tinker
