AI Nutrition & Recipe Intelligence Engine.

Destra (specifically the core intelligence engine in destra/utils) is an automated AI and serverless data pipeline designed to parse, harmonize, normalize, and enrich complex nutritional data from raw web recipes and unstructured culinary content.

Engineering Challenge: Real-world culinary data is inherently noisy—featuring non-standard ingredient descriptions, compound fractions, modifier adjectives ('finely diced roasted red pepper'), implicit cooking transformations (oil absorption during frying), and ambiguous portion units. Traditional regex-based parsers fail on edge cases, while pure LLM solutions are too slow and expensive at scale.

The Multi-Stage Pipeline Architecture:

1. Scraper & Harmonizer: Extracts structured JSON-LD/Microdata and cleans instructional content with automated fallbacks.
2. Hybrid Deterministic & LLM Parser: Combines fast deterministic quantity/unit regex logic with targeted LLM fallback for non-standard items.
3. OpenSearch Serverless Retrieval: Queries indexed USDA Food Data Central (FDC) Foundation and Branded food databases using custom query term expansions, hit ranking algorithms, and modifier score penalties.
4. Semantic Identity Scoring & Normalization: Evaluates match confidence using semantic identity algorithms, unit/gram conversions, and embedded portion unit matching.
5. Nutrition Pass/Fail Parking Lot & Review Queue: Automated benchmarking system that evaluates macro calculations against source benchmarks, enqueuing low-confidence items into a human-in-the-loop review workflow.

Infrastructure & Performance: Built with AWS Amplify Gen 2, DynamoDB TTL caching, AppSync GraphQL/REST handlers, and Bun test harnesses. The pipeline achieves sub-second cached responses, high-accuracy normalization, and zero memory leaks across Lambda lifetimes.