Case studiesLoafy

1,500+ questions answered from one book, at under 1 cent each

Loafy answers bakers from my open-source book, in their language, and names the chapter: 1,500+ questions in seven weeks, at under 1 cent of model cost each.

Role
Founder
Period
Since 2025
Status
In production since January 2026
The Loafy iPhone app: a baker asks why the dough is sticky, and Loafy answers from the book, naming the chapter
  • 1,500+ questions answered since September 2026
  • 690+ registered bakers in 70+ countries
  • <1 ¢ model cost per answer
  • 750+ bakes logged in the journal

In 30 seconds

The problem

Bakers get stuck mid-bake, and the internet gives them ten contradicting answers. The right one is in my book, but nobody searches a book with dough on their hands.

What I built

Ask Loafy: questions in the baker's own words, answered with the whole book in context, every answer naming its chapter. On the web and on iPhone.

Outcome

1,500+ questions answered in seven weeks, 690+ registered bakers in 70+ countries, under 1 cent of model cost per answer.

For you

The path of every AI proof of concept: success defined up front, options tested blind on your own material, the cost of every answer on record.

The challenge

Sourdough bakers get stuck in the middle of a bake: the dough is slack, the starter will not rise, the crumb is dense. Search engines and forums give them ten contradicting answers. The right answer is usually in The Sourdough Framework, my open-source book, but nobody searches a whole book with dough on their hands.

Loafy started in 2025 as a landing page that linked to an AI notebook holding the book. The question for 2026: can an AI answer from the book itself, reliably, in the baker’s own language, at a cost low enough for a free product?

What I built

Loafy, a free baking companion on the web and on iPhone and iPad. Its core is Ask Loafy: a baker asks in their own words; Loafy answers from the book and names the chapter each answer comes from. When the book does not cover something, Loafy says so in its first sentence.

Before building, I wrote down what success means: questions per day, sign-ups from the daily limit, the thumbs-up rate and the cost per day, checked after two and after four weeks. Then I measured instead of assuming. In a blind evaluation of 68 questions, graded against the book by an AI judge, the existing retrieval pipeline scored 7.34 out of 10 and missed answers the book contains. The whole book in context, in the pipeline that went to production, scored 9.65. The cause was retrieval, not the model, so retrieval went. The model itself was chosen the same way: 17 models graded blind against the book, the winner picked on real questions from real bakers, and switched twice in six weeks when the evidence changed.

A Loafy answer to a crumb photo: what it sees, the position on the crumb map from the book, marked as provisional, and what to do next
The crumb check: a second opinion on a photo of the cut loaf, placed on the crumb map from the book, and honest about how sure it is.

Around Ask Loafy sit the tools a baker needs on the day: the crumb check for a photo of the cut loaf, a bake planner that works backwards from when you want bread, guided first bakes, bread and pizza calculators, and a bake journal that Loafy reads back to the baker next time. Guardrails were set before launch: daily question quotas and hard daily cost caps.

The result

Since Ask Loafy went live in September 2026, it has answered 1,500+ questions from bakers, at under 1 cent of model cost per answer. Since mid-October every answer names the chapter it comes from. 690+ bakers in 70+ countries have registered and logged 750+ bakes in their journals, and the iPhone and iPad app has been in the App Store since October 2026.

That is the path an AI proof of concept takes in your company: success defined before the first line of code, options tested blind on your own material, a system in production within weeks, and the cost of every answer on record.

For the Head of IT

  • Stack: Ruby on Rails with PostgreSQL and Hotwire on the web, a native SwiftUI app on iOS, one API for both with streamed answers, self-hosted in the EU. Models are called through one gateway, so the model is a configuration choice, not a rewrite.
  • Evidence on record: every answer is stored with its model, tokens, cost and latency; the median full answer takes about 6 seconds. 7,500+ automated tests run on every change to the web app, 1,300+ on the iOS app.
  • Verify it yourself: ask Loafy a question at loafy.ai/ask, and read the book it answers from at github.com/hendricius/the-sourdough-framework.

What could AI deliver in your company?

Book an intro call (30 min) Or start with the AI proof of concept package →

Free, no slides, no obligation: a direct conversation about your situation.