An Evaluated End-to-End Opinion Intelligence Platform: Transparent Multi-Engine Sentiment Analysis, Hybrid Routing, and Evidence-Oriented Workflows

Sentiment components are often evaluated separately from the software workflow that preserves their inputs, provenance, and outputs. This applied AI systems study evaluates an opinion-intelligence platform integrating Next.js/React, typed services and repositories, a PostgreSQL/Supabase persistence design, and FastAPI inference. Historical evidence is preserved from a seed-20260622 stratified sample of 5,000 airline records, on which RoBERTa achieved 0.8098 accuracy and 0.7587 macro F1. A later study adds Financial PhraseBank and sentence-level SST-5 collapsed into three classes; TweetEval remains reference-only because of the pretrained model’s provenance. A lightweight logistic router uses VADER and textual features, fits only development data, and freezes a validation-selected threshold before domain-specific test evaluation. RoBERTa test macro F1 was 0.7495, 0.5586, and 0.6278 on airline, finance, and SST, respectively. The router reduced transformer use, but its test macro-F1 loss exceeded the 0.02 validation tolerance in all three domains. Ten post-warmup repetitions per batch and serving configuration supplement bootstrap intervals, paired correctness tests, and calibration diagnostics. Batching changed throughput without changing predictions. Local checks exposed ownership, denied-write-response, and print-layout defects; repairs passed isolated regression checks, while hosted verification remains incomplete. Human error coding and user validation remain incomplete. The contribution is transparent system integration and empirical serving evaluation, not a new sentiment model or routing algorithm.