Khrystyna DidukhFree call
← All cases

AI nutrition tracker

Sila

Log a meal by voice, text or photo — and correct it in plain language. A hybrid model-routing layer cut the AI cost per entry many times over without losing accuracy.

Role
Product, design and build — end to end, with AI
Made with
Built with Claude and Lovable

01Context

Sila was built around the way people actually talk about food: say it, photograph it, then amend it — "I didn't finish the bread" — and the entry updates itself.

02The problem

Nutrition apps fail on friction. Typing every ingredient is work, and every correction means starting the entry again.

03What was built

  • 01Meal logging by voice, free text or photo.
  • 02Natural-language corrections applied to an existing entry, not a new one.
  • 03Leftover subtraction from a second photo of the plate.
  • 04A personal food memory: the user's own recurring dishes are recognised and reused.
  • 0514 interface languages.
  • 06Hybrid AI model routing — cheap models handle the common path, stronger models are called only where accuracy demands it.
  • 07Credit-based pricing so cost per user tracks actual usage.

04Automation

How the flow runs — input, AI, the human decision point, and what comes out.

01Voice, text or photoraw meal input
AIRouter picks the modelcheap path first, escalate on doubt
HUMANUser corrects in words"I didn't finish the bread"
04Adjusted entry + food memorynutrition log updated
Fig. 01 — Meal entry and correction loop

05Stack

  • Claude
  • Lovable
  • Multimodal LLMs (vision + speech)
  • Model routing layer
  • React / TypeScript
  • Postgres
  • Credit billing

06Numbers

Interface languages
14
AI cost per entry
Reduced many times over via model routing
Exact cost reduction
[add ×]

Editable placeholder

Entry accuracy
[add %]

Editable placeholder

Active users
[add number]

Editable placeholder

07Screens

From the live app.

Nutrition tracker Today screen on a phone: calorie ring showing 1,279 of 1,900 kcal, protein, fat and carb bars, the meal list, and the capture bar with barcode, voice, text and photo buttons.

Today: one ring, the whole day

"Confirm entry" sheet after typing "salmon with rice and salad": salmon fillet 140 g, cooked rice 150 g and fresh salad, each editable, with a field to correct the entry in plain words.

Say it, check it, save it

History screen on a phone: September calendar with daily calorie totals, and the selected day's entries below.

History: every day, at a glance

Weight screen on a phone: 63.9 kg, −0.9 kg since start, a trend chart and the daily weight log.

Weight: the trend, not the noise

Want this kind of result in your team?

I teach teams to build on their own work the way I built Sila. Start with a free call.

Book a discovery call

20–30 minutes · No commitment · Get a reply within 24 hours

Next case