We were building an AI-powered math app and were seven weeks from MVP when I discovered all the competitor apps—with the exception of Khanmigo—used image capture as its primary ingestion method. The Engineering team didn’t have time to add a custom loader to the roadmap, so I built one over that weekend using Python, Streamlit, and the Mathpix API.
One challenge was that Mathpix returns LaTeX (used for math and science markup) separately from plain text, which created problems for word problems that needed both formats integrated. To solve this, I wrote custom logic to align the LaTeX output by tracking the bottom-left corner of each character and reconstructing the full problem in order.
This became the approach I’ve taken to building POCs. I prefer Streamlit over Figma because it’s not so scripted, and your stakeholders can access it early on. Furthermore, you aren’t stymied by access issues if you use publicly available data that’s similar to the data that will be ingested—or synthetic data.
Note: This was earlier in my journey as a coder. I now follow a pretty strict structure and format for my code and comments.