Raw Reviews (ingestion & preprocessing)

Shared Transformer Encoder (BERT/ DistilBERT/ RoBERTa)

LangGraph
Orchestrator
(DAG execution & state)

Agent 1: Feature Discovery — Multi-label classifier

Agent 2: Feature Sentiment — Polarity (+/0/-) & confidence

Calibration (sigmoid α,β; optional T)

Agent 3: Feature-Grounded Summarizer (LLM constrained to features)

Outputs:
Feature list + per-feature sentiment & ratings • Faithful summary

Rating Aggregator — Uniform/ Learned weights

Artifacts & Logs (checkpoints, metrics, reliability diagrams)