Sourdough Intelligence

Background

This project started in 2018, before LLMs were part of anyone’s toolkit.

The original goal was simple: figure out which sourdough recipe gave a first-time baker the highest probability of success on the first attempt. To do this, I built a two-stage model:

1. Multiple linear regression across key recipe variables: hydration, fermentation time, flour type, inoculation rate, fold count, and bulk temperature.
2. Sentiment analysis on YouTube comments (IBM Watson NLP) to cross-reference which recipes correlated most with beginner success, and which ones consistently produced frustration or failure.

The output was a ranked shortlist of top 3 beginner recipes. I picked one, baked it, and it worked first try.

No LLMs. No fancy stack. Just R, IBM Watson NLP, and a probably unhealthy amount of time reading bread forums.

The product

Crumb turns that research into an interactive browser app. Instead of asking a baker to follow a fixed recipe schedule, it asks:

• What are you baking?
• When do you want it out of the oven?
• How much active handling time do you have?
• How warm is your kitchen?
• What flour and skill level are we designing around?

The app then generates a practical formula, ingredient list, and timestamped schedule that fit those constraints.

Product highlights

• Multi-step bake wizard for bread type, timing, active-time budget, kitchen temperature, flour, skill level, loaf size
• Seven bread archetypes: sandwich loaf, toast loaf, country loaf, rolls, pizza dough, focaccia, baguette
• Recipe formula generation using baker’s percentages
• Schedule engine that works backwards from the target finish time
• Active-step planning with quiet-hours support: folds and shaping don’t land during sleep or work blocks
• Temperature-aware bulk fermentation estimate
• Warnings for unrealistic timelines, risky hydration, hot kitchens
• Printable plan and PNG export, light and dark themes, dependency-light static deployment

Try it live →