Maths apps like Khan Academy Kids or Prodigy, and photo solvers like Photomath, either use invented problems or solve problems a child already has. Teachers know "real world" maths works, but writing problems about each child's own street is impossible by hand. Street Math turns a photo of the child's actual street into a graded set of problems (count these windows, use this shadow, price this stall in your own currency), checked step by step.
The data it runs on
Photos the child (or a parent) takes, plus optional prices typed from a real stall. A vision model detects countable objects (windows, wheels, fruit), finds a reference height (a standard door is about 2 m, a lamp post's height can be entered) and measures shadow lengths in the image. Limits: counts in busy photos can be off by one or two, and shadow maths needs a sunny photo with a visible base. Fallback: the child confirms every count with a tap before problems are generated ("Is it 12 windows?"), and if there is no clear shadow, measurement problems switch to estimation using doors and storeys. Photos are processed and deleted; only the detected objects are kept.
1. The mission
Make every child's own street their maths textbook. When the numbers are about the fruit stall your aunt runs or the windows you walk past every day, maths stops being somebody else's story.
2. Why people can't resist it
It is about me. Children care about their own street far more than about "Sam has 7 apples".
Treasure hunt energy. Going outside to find a shadow or count windows feels like a game, not homework.
Visible progress. Topic bars and a street streak show growth to children, parents and teachers at once.
3. How it works for real
Capture: phone camera, or a parent's phone for younger children. Location rounded to city level for currency and latitude only.
See: a vision model (Gemini Flash vision or a GPT vision model via OpenRouter) returns a JSON list of objects with bounding boxes, counts and shadow endpoints. A small open detector (YOLO-class) double-checks counts on device.
Confirm: the child taps to confirm counts and types or accepts prices.
Write: a frontier LLM (Claude or GPT) writes problems from the confirmed scene graph, mapped to curriculum objectives by age (counting, money and change, multiplication, ratio, fractions).
Check: answers are computed in code from the scene graph, never by the LLM. Hints come in two levels plus a "show me" highlight on the photo.
Teach: the teacher view turns any photo into a printable worksheet with an answer key, and each pupil's street becomes their own version of the same skills.
4. Cost per use
Pennies per worksheet: one vision call (about $0.002) plus one problem-writing call (about $0.005). Hints and checks run in code.
Free: 3 streets a week for families. Family ($3 a month): unlimited streets, progress reports. School ($1 per pupil per year): class sets, worksheets, curriculum mapping.
5. Growth loop
A child shares "I solved my street" with a friend, who wants to try their own street. Teachers set "photograph your street" as homework, so a whole class signs up at once, and parents see the progress card.
6. Risks and guardrails
Child safety: faces and number plates are blurred before anything is stored; no house numbers or street names leave the device.
Location: only the city is used (for currency and latitude), never GPS coordinates.
Wrong maths: answers come from code, and counts are confirmed by the child.
Safety outdoors: prompts say "with an adult" and never ask children to approach strangers or traffic.
7. MVP build order
Photo upload and a vision prompt that returns windows, people-free objects, shadows and price cards as JSON.
Confirm screen, then problem templates for 3 age bands computed in code.