Show Dev: An AI Photo Solver I've Been Working On

Show Dev: An AI Photo Solver I've Been Working On

I have been experimenting with a small AI study tool that starts with a photo instead of a typed prompt.

The idea is simple: a student takes a picture of a homework problem, the system extracts the question, routes it through a few model paths, and returns step-by-step explanations that can be compared.

👉 Download Now from the App Store: https://apps.apple.com/us/app/ai-snapsolve-homework-solver/id6763911277
App Store Search: AI SnapSolve

This post is less of a launch announcement and more of a development note: what felt interesting, what was harder than expected, and why the "photo first" interface changes the product shape.

Why start with a photo?

Typing math or science questions into a chat box is still surprisingly awkward.

Fractions, diagrams, units, handwritten notes, and multi-part questions all add friction. A photo lowers that friction. It also preserves context that students often forget to type: the diagram on the side, the wording of part B, the previous line of work, or the exact format their teacher used.

That means the first challenge is not "solve this problem." It is "understand what problem is being shown."

The workflow I built around it

The current flow has a few stages:

• read the image and identify the question text
• keep visual context such as diagrams or formatting where it matters
• classify the subject area
• send the problem to more than one solving path
• return explanations that show reasoning, not only the final answer

I found that this structure makes the product feel less like a search box and more like a review surface. The user is not only asking for an answer. They are checking whether the system understood the problem correctly and whether the reasoning makes sense.

Why use more than one model path?

One thing I wanted to avoid was a single confident answer with no way to compare it.

For homework-style questions, multiple approaches can be useful. A quadratic might be solved by factoring, using the quadratic formula, or reasoning from a graph. A physics problem might be modeled with different assumptions before arriving at the same result.

So the tool uses multiple solving engines and presents the outputs as alternative paths. This is not a perfect guarantee of correctness, but it gives the student more to inspect:

• Do the answers agree?
• Which method is easier to follow?
• Did one path skip an important step?
• Is there a mismatch between the image and the extracted problem?

That comparison layer has become one of the most useful parts of the experiment.

The part that surprised me

The hard part is not always the model answer.

Often, the difficult part is the messy middle: recognizing handwriting, preserving notation, detecting whether two photos belong to the same question, and deciding when a diagram is essential rather than decorative.

Multi-image upload also changes the logic. A worksheet can span several pages, and part 3 may depend on part 1. Treating every image as an isolated prompt is easy, but it is not how real homework works. The system needs to merge the images into one context without inventing missing information.

Keeping the output useful for learning

I have been trying to keep the explanations focused on understanding rather than speed alone.

A fast answer is convenient, but a useful study answer should expose the structure of the problem:

• what information was identified
• which formula or concept applies
• why a particular step follows from the previous one
• where a student should double-check their own work

This is especially important when AI is used by students. The interface should encourage review, comparison, and verification. It should not make the final answer feel like the only thing that matters.

What I would still improve

There is plenty left to refine.

I would like better handling for ambiguous handwriting, clearer confidence signals when the image is low quality, and stronger ways to flag when two model paths disagree. I also think the answer comparison UI can become more educational by making differences between approaches easier to scan.

Another area is feedback. If a student says "I don't understand step 3," the tool should be able to zoom into that step without restarting the whole solving process.

Closing thought

The interesting part of this project is not just that AI can solve a problem from a photo. It is that the photo can become the start of a learning workflow: read, interpret, compare, explain, and revise.

That feels like a more useful direction than treating homework help as a one-shot answer machine.