Restaurant-BERT-Triage

Restaurant-BERT: Low-Latency Intent Classification Engine

Overview

This model is a fine-tuned DistilBERT architecture (distilbert-base-uncased, ~66M parameters) designed to serve as a deterministic, low-latency intent classifier for restaurant customer queries. It acts as a triage layer to route messages before they reach expensive Generative LLM pipelines.

Routing Logic & Taxonomy

The model maps incoming text payloads to four deterministic intents:

• order: Queries related to placing or modifying an order.
• inquiry: General questions (e.g., operating hours, location, menu items).
• greeting: Standard conversational openers.
• support: Complex issues requiring human intervention or advanced LLM processing.

Performance Metrics & Limitations

• Baseline Accuracy: 94.50%
• Average Confidence: 80.57%

Evaluation Note: The current weights were evaluated against a synthetic holdout test set. Because both the training and testing data were generated synthetically, real-world inference confidence on organic human phrasing is expected to exhibit a distribution shift. Retraining on organic chat logs is recommended for full production deployment.

Full System Implementation

This model is designed to be hosted locally via FastAPI and orchestrated using n8n webhook routers to achieve ~50ms inference latency with zero API costs.

For the complete orchestration architecture, FastAPI inference scripts, and n8n workflow exports, please refer to the main project repository: https://github.com/Belomoh6/Restaurant-BERT-Triage