HowSure

Inspiration

I already used a personal food tracker and wanted more useful information with less mental effort. My vision was a simple dashboard where tapping a number reveals its evidence, assumptions, and limitations.

HowSure turns that idea into a standalone app: show the estimate, explain its uncertainty, and show what information could improve it.

What it does

HowSure supports manual meal logging, basic macros, and calorie estimates with an estimate score and two model-based uncertainty ranges. Each meal records how its numbers were obtained, such as a nutrition label, published menu, photo estimate, or description.

The fictional demo starts at 2,280 kcal with an estimate score of 83. Updating one meal with its nutrition-label information changes the total to 2,380 kcal, raises the score to 92, and narrows the range. A before-and-after preview shows the effect before saving.

Missing nutrition remains unknown rather than becoming zero. Meals follow waking days, so an after-midnight snack stays with the day the person was awake for. Sleep duration, a seven-night sleep average, and local saving are also included.

How we built it

I directed the product decisions and used Claude Code with the Devpost Learn Skill Pack for planning and AI-assisted implementation. The app uses React, TypeScript, and Vite, with a separate calculation core and browser-local storage.

Expected calculation results were computed independently in Python. Verification included 240 unit tests, 76 browser tests, and a clean type check. Playwright recorded the working app for the demo.

The implementation started from an empty folder. Ideas and starting error parameters came from my earlier personal tracker and are disclosed in the README; its code was not carried over.

Challenges we ran into

The hardest issue was defining an estimate score without implying scientifically validated accuracy. The source review did not establish the exact starting error parameters, so every parameter remains labelled as an assumption. The displayed ranges are model outputs, not guaranteed bounds or empirically calibrated coverage.

Time constraints required freezing the verified core and documenting unfinished features. Limited memory during recording was handled by capturing short scenes separately and joining them afterward.

Accomplishments that we're proud of

The central interaction works end to end: inspect uncertainty, improve a meal's evidence, preview the effect, save, and see the day's calculation change.

The app also distinguishes unknown values from zero, preserves entries after reload, and handles waking days across midnight. Its explanations make the calculation's assumptions visible.

What we learned

A correctly implemented calculation is different from an empirically validated estimate. Independent test values help verify the implementation, while source labels explain what still rests on assumptions.

The project also demonstrated the value of turning a broad vision into observable requirements and verified build slices, with unfinished work recorded explicitly.

What's next for HowSure

First, finish demo reset, first-wake regrouping previews, and historical sleep/wake editing with warnings before meals move between days.

The broader roadmap includes weekly calorie views, weight trends, food lookup, Claude-assisted logging and research, and a carefully defined gut-friendly diet score. These are future features, not features of the submitted build.

Built With

• claude-code
• devpost-learn-skill-pack
• localstorage
• playwright
• react
• typescript
• vite
• vitest
• zod