Developed a full-stack nutrition estimator application that uses Gemini (LLM) to extract ingredients and serving data from user input, and then matches these ingredients to a locally stored nutrition database (Excel) to compute estimated nutritional values per serving. The project combines natural language understanding, alias matching using LLM, and fuzzy string comparison to accurately calculate nutrition information for Indian dishes and user-defined meals.
1- Gemini-Powered Ingredient Extraction:
- Used Gemini API to parse the user input (dish name or description) into structured data: dish name, ingredients used, intent, serving size, and assumptions.
2- Alias Generation Using Gemini:
- For each ingredient, a secondary Gemini call was made to generate up to 4 common aliases or alternate names, including English terms. This was to improve matching accuracy against your Excel dataset.
3- Excel Dataset Integration:
- Parsed a custom Excel file (IFCT 2017 based) using xlsx, extracting and cleaning food items with fields like
energy_kcal, protein_g, carb_g, fibre_g, etc.
4- Fuzzy Matching with string-similarity:
- Compared user-provided (and alias-enhanced) ingredient names with food names from the Excel dataset using string similarity, with a confidence threshold for matching.
5- Nutrition Calculation Logic:
- Computed aggregate nutrition per serving based on matched food items and returned a complete breakdown of macronutrients.
6- Express API Endpoint:
- Built a dedicated NutritionEstimator endpoint in Express.js that ties all components together and returns a detailed JSON response.
1. Challenge: Inconsistent Ingredient Naming
-
Problem: Users input ingredients with regional or colloquial names (e.g., rajma, baingan, chana), while the Excel dataset used scientific/English names.
-
Solution: Made a separate Gemini API call to generate multiple aliases per ingredient. This expanded the matching vocabulary and drastically improved hit rate against the database.
2. Challenge: Low Matching Accuracy
-
Problem: Direct string matching failed due to slight spelling variations or different word orders.
-
Solution: Used string-similarity to find the best match for each alias, with a confidence threshold (e.g., >0.8) to reduce false positives.
3. Challenge: Parsing Gemini Responses
-
Problem: Gemini sometimes returned explanations or wrapped JSON in text, making parsing difficult.
-
Solution: Extracted JSON content from within the full response using string slicing and error-handled parsing to prevent crashes.
4. Challenge: Handling Missing Data
-
Problem: Some ingredients had no direct match in the dataset.
-
Solution: Returned a fallback using the original ingredient name or skipped low-confidence matches to maintain data integrity.
5. Challenge: Working with Excel Data in ESM
-
Problem: Using __dirname in ES modules for path resolution was non-trivial.
-
Solution: Used fileURLToPath and path.dirname to resolve paths in an ESM-compatible way.
- Node.js, Express.js
- Google Gemini API
- Excel Parsing with xlsx
- String Matching with string-similarity
- JavaScript (ESM modules)
- Data normalization and transformation