Thai Review Sentiment Intelligence

Overview

A Thai sentiment workflow connecting inference, confidence-aware routing, human review, monitoring, and governance reports.

Problem

Sentiment predictions need transparent confidence handling, review queues, and monitoring when uncertain Thai-language inputs affect downstream decisions.

Solution

Built a Thai NLP workflow with classification, confidence routing, human review, explainability metadata, monitoring, and an active-learning queue.

Technical Decisions

• A deterministic classical ML baseline keeps behavior inspectable.
• Low-confidence output is routed for review instead of forced into automation.

Outcome

The project exposes model behavior and governance evidence instead of presenting classification output as an unquestioned decision.

Key Features

• Confidence routing sends uncertain predictions to human review.
• Explainability metadata accompanies model output.
• Monitoring and active-learning queues turn feedback into governance evidence.

Architecture

1. 01

   Thai review

2. 02

   Text preprocessing

3. 03

   Classifier

4. 04

   Confidence routing

5. 05

   Human review

6. 06

   Monitoring

7. 07

   Active-learning queue

Tech Stack

• Python
• scikit-learn
• FastAPI
• React
• Thai NLP
• Monitoring

Verification

• Model evaluation artifacts
• Monitoring demo
• Governance and active-learning reports

Security & Privacy

• The public workflow does not require private customer review data.
• Human review remains part of the decision path for uncertain predictions.

Limitations

• Sentiment labels simplify context such as sarcasm, mixed sentiment, and domain-specific language.
• The portfolio links to repository evidence without claiming a public deployment.

Future Improvements

• Expand evaluation across domains and language variation.
• Calibrate confidence thresholds with reviewer feedback.