The Semantic Flavor Graph for High-Volume Platforms.
Generate context-aware upsells in <50ms. Built on Sentence Transformers and Neo4j—not generic collaborative filtering. Solves the cold-start problem instantly.
"meta": {
"latency_ms": 42,
"model": "v2.1.0-hybrid"
},
"recommendation": {
"item": "Truffle & Parmesan Fries",
"scores": {
"semantic_sim": 0.89,
"chem_affinity": 0.98
},
"reasoning": {
"driver": "flavor_bridge",
"notes": "High molecular overlap via glutamates."
}
}We treat flavor as a high-dimensional vector space.
Standard collaborative filtering fails on new items. We analyze the food itself.
1. Ingestion & Vectorization
We use all-MiniLM-L6-v2 transformers to convert menu text into 384-dimensional vector embeddings. The model understands "Guanciale" contextually without purchase history.
2. The Flavor Graph
Vectors are mapped against a Neo4j Graph Database containing validated chemical pairings. We traverse edges to find non-obvious, high-margin pairings.
3. < 50ms Inference
Pre-computed embeddings stored in memory. The final output is a weighted score of semantic similarity + chemical affinity.
See the Lift.
Moving from "Most Popular" to "Semantic Compatibility" drives a 4x revenue lift on attachments.
Built by Data Scientists, Not Marketers.
Culinary Compass is an engineering-first operation. Led by an MSc Statistician. We don't sell "magic AI." We sell clean, containerized Python code that integrates into your checkout flow in minutes.
Start Integration
Developer Sandbox
- 50 API calls / day
- Full Documentation
- Community Support
Early Adopter
- 25,000 API calls / month
- < 50ms Latency SLA
- Direct Engineer Support
- Custom Model Tuning
Request API Access
Supported By
