How Eatomate Works

Receipt printers create predictable errors. We use research-backed confusion matrices that understand certain character pairs are commonly confused by OCR. This includes both single-character substitutions and multi-character segmentation errors where two characters are misread as one (or vice versa). Each confusion has a likelihood score.

Instead of treating all typos equally (standard Levenshtein distance), we weight matches by how likely the OCR error is. A common confusion costs less than a rare one, leading to better matches on real-world receipt data.

When you map a receipt item to a barcode once, the system remembers it with a confidence score. Repeated mappings increase confidence. The cache learns your shopping patterns at specific stores. By week 12, automated matching handles 96% of receipt items.

Your search text is matched against three databases (Barcode Database 2M+, Recipe Database 50K+, Alternative Names 100K+). If no exact match exists, Mistral AI generates a personalised recipe variant with full nutrition data, added to the database automatically.

Each food item has multiple aliases (e.g., "Shepherd's Pie" → ["Cottage Pie", "Mince and Potato Pie", "Meat and Tatties", ...]). Your search text is matched against ALL aliases in our 100K+ recipe database, ensuring you find the right item even with different naming conventions.

When no match is found (>80% confidence threshold), Mistral AI generates a complete recipe variant with:

• Ingredient list with quantities and classes
• Full nutrition profile per 100g
• Confidence scoring and usage tracking

Variants are automatically added to the database with no manual review, creating a personalized recipe collection.

When you finish an ingredient (milk carton empty), we use the equation: consumed = initial + purchased - waste - final. All quantities tracked in grams with uncertainty (e.g., 250g ± 15g). Gaussian error propagation combines uncertainties mathematically, locking historical meals at high accuracy.

Uncertainty reduces with more observations. Initial prediction: 250g ± 37.5g (±15%). After 10 observations: 250g ± 11.8g (σ/√10). After 30 observations: 250g ± 6.8g (σ/√30). Bayesian combination of multiple measurements (weight entries + pantry reconciliation + historical learning) further reduces uncertainty. This is how we achieve ~95% accuracy by Week 4.

System learns your ingredient ratios automatically using weight (grams), not volume. Your bolognese might use 12g oil/100g while the generic database says 5g/100g. Reconciliation discovers this and updates recipe accuracy using exponential moving average with α=0.3 learning rate. Weight-based tracking eliminates density ambiguity (ml→g conversion errors).

All displayed values include 2 standard deviation (2σ) error bars, representing 95% confidence intervals. This means we're 95% confident the true value lies within the shown range. For example, "250g ± 15g" means we're 95% certain the actual portion was between 235g and 265g.

Unlike other apps that show false precision ("exactly 247.3 calories"), we show honest uncertainty ranges. Early weeks have wider error bars (±15-20%) which narrow over time (±5-8% by Week 4) as the system learns your specific patterns and recipes.

Using 95% CI (2σ) is the gold standard in scientific research and clinical trials. This approach allows for honest comparison with lab-based methods like doubly labeled water (DLW), which itself has ±3-8% measurement uncertainty. Our ~95% accuracy claim is measured against DLW ground truth with proper statistical methodology.