jaay2/bawago-universal-model

BawaGo Universal Hierarchical AI Agent Model

A hierarchical semantic ensemble model trained for BawaGo peer-to-peer capacity matching, covering all 7 verticals (Baggage, SafeVault Escrow, Rides, Crave Home Food, Host Homestays, Wish Shopping, Fresh Farm Produce) across 130 universal intents.

Model Summary

• Architecture: Two-Stage Hierarchical Classifier + Domain Keyword Ensemble
  • Stage 1: 13-Class Vertical Router (99.75% accuracy)
  • Stage 2: 130-Class Granular Intent Classifier with sublinear TF-IDF (99.55% accuracy)
  • Stage 3: Context Agreement & Keyword Safety Layer
• Dataset Size: 1,997 curated multi-vertical query-intent pairs
• Inference Latency: ~1.2 ms on standard CPU
• Framework: scikit-learn >= 1.4.0, joblib

Files Included

• trained_agent_model.joblib: Serialized 2-stage Hierarchical Pipeline (12.4 MB)
• universal_kb.json: Knowledge base containing 130 non-technical, conversational policies
• intent_answers.json: Intent-to-response mapping with action checklists
• dataset.json: Full 1,997-sample training dataset
• model_metadata.json: Model configuration and training report