Inspiration
A recipe told over the phone often assumes you know which bowl, how full and how many people it feeds. A transcript preserves those words, but the missing measurements remain missing.
MeasureBack focuses on that gap. It asks for a measured calibration, reads quantities back and keeps a question open when the cook cannot supply an answer. The cook can use an ordinary phone. The person preparing the dish gets a recipe with its source quotations beside it.
What it does
The walkthrough starts with two bowls of rice for four people. Until the bowl is measured, the rice quantity stays unresolved. A 250 ml calibration makes that 750 ml for six people. Later, the cook corrects water from 1000 ml to 900 ml. MeasureBack keeps the correction and scales the updated amount to 1350 ml.
Select an ingredient to inspect the exact words behind it. Change the serving count without losing the selected quotation. Salt remains to taste, fractional pieces require a decision, and the stated simmer and rest times remain unchanged.
The local application also prepares an approved CALL E interview and retrieves its structured result. English, Hindi and Tamil are available as initial instructions, with fixed or adaptive language mode. Adaptation is a best effort instruction to the provider, not a verified automatic language detection feature.
How I built it
CALL E handles the phone conversation and structured extraction. Python checks recipe structure, source quotations, corrections and unit compatibility, then calculates amounts using exact fractions. The browser displays those calculations beside their evidence. It does not invent a bowl size or convert volume to weight without a measured calibration.
The interface borrows from a kitchen scale: one selected measurement, a large numeric display and a recipe ledger. The public walkthrough is a static export of three authored stages calculated by the same Python engine. Private imports and calling stay in the local application.
Challenges
A valid JSON response can still misread a correction or quote the wrong speaker. MeasureBack checks extracted cook turns against the provider transcript in order. The source text remains visible for human review because exact quotation matching cannot establish semantic correctness.
Consent needs the same care. Returned recipes require transcript anchored capture and sharing permission, followed by a review of the full conversation. A persistent pending call reference prevents an uncertain submission from turning into an accidental new interview after a reload.
What is verified
The project passes 140 Python tests, 20 local browser checks and 10 independent checks of the public deployment. A local HTTP provider exercise verifies submission, result handling and duplicate prevention without dialing anyone. The walkthrough and video show an authored conversation running through the real application. No live cook interview, language quality result or real user outcome is claimed.
Voice recipe capture already exists. MeasureBack's specific focus is resolving household measurements through clarification, retaining corrections and exposing the evidence behind each calculation.
What's next
Test the interview with consenting cooks and measure whether it asks useful clarification questions in each supported language. Recipe quality, safe cooking technique and live language behavior still need that validation.
Log in or sign up for Devpost to join the conversation.