/ ENGINEERING SYSTEMS
← Projects

RoBERTa · NLP · Product Signals

Customer Sentiment & Product Intelligence

2026RoBERTa

RoBERTa-based sentiment scoring across customer channels with category-level reporting and statistically significant change alerts.

THE PROBLEM

Why this system existed

Aggregate sentiment is less useful without product attribution, trend monitoring, and a path from signal to action.

OUTCOME

What changed

Processed 50,000+ records daily at 89.1% accuracy and surfaced signals that led to three product corrective actions.

REFERENCE ARCHITECTURE

Controls around the model

01Classify
02Route
03Retrieve
04Rerank
05Generate
06Validate
07Observe

DECISIONS

Trade-offs considered

Five-class sentiment scale
NER-based product attribution
Rolling-mean anomaly alerts
Automated weekly reporting

FAILURE CASE

What did not work

Lexicon baselines missed domain nuance; a RoBERTa model and category attribution produced more actionable product-level signals.