Helping moderators make consistent, evidence-based decisions—without relying on opaque model outputs.

🔗 Live Demo: https://ai-review-moderation-2-hdqedtqcmlbagcmiewmebt.streamlit.app

📂 GitHub: https://github.com/itingtseng/ai-review-moderation-2

🎯 Problem & Motivation

Moderators often review hundreds of borderline cases per day, under time pressure and policy ambiguity.

✅ What’s New vs v1.0

Based on observed limitations in classifier-only moderation systems, v2.0 introduces…

v1.0 v2.0
Classifier-only decisions Hybrid (rules + neighbors)
Hard-coded keyword list Data-driven log-odds n-gram
No fallback Graceful degrade to rules
No queue Moderator panel
Hard-coded thresholds UI sliders

🧠 Hybrid Decision Design

final_score = α·rule_score + β·neighbor_conf

♻️ Governability

🧯 Fallback Logic

If FAISS index unavailable:

neighbor_conf = 0

→ still deterministic

👉Key Reviewer Experience Decisions

Designed to reduce cognitive load and support consistent human judgment under uncertainty.

🔍 Explainability

🧑‍⚖️ Moderator Queue

📊 Insights Examples

🔥 Product Impact

🧰 Tech Stack

Layer / Function Tool / Library Reason
Frontend UI Streamlit Fast iteration, moderator-friendly UI
Backend API FastAPI Async, low-latency scoring endpoints
Embeddings SentenceTransformer (all-MiniLM-L6-v2) Lightweight semantic understanding
Vector Search FAISS Efficient Top-K similarity retrieval
Rules Engine YAML (rules.yml) Versionable governance config
Explainability Regex / token stats Transparent evidence for moderators
Fallback Logic Custom degrade handler Maintains determinism if index missing
Configuration .env / environment vars Deployment toggles (thresholds, modes)
Logging / Audit Python logging Traceable moderation decisions
Visualization Streamlit components Inline evidence panels
Data Processing Pandas / scikit-learn Token stats, n-gram extraction
Deployment Streamlit Cloud Zero-ops hosting

📦 Architecture

Shown for implementation reference

flowchart TD
    A[User Review] --> B[Preprocessing]

    B --> C[SentenceTransformer Embedding]
    C --> D[FAISS Top-K Neighbors]

    B --> E[Rule Engine rules.yml]
    E --> F[Rule Score]

    D --> G[Neighbor Confidence]

    F --> H[Hybrid Decision α·rule + β·neighbor]
    G --> H

    H --> I[Risk Tier: HIGH MED LOW]

    I --> J[Moderator Queue Panel]
    J --> K[Human Verdict Logging]

    H --> L{FAISS available?}
    L -->|No| M[Graceful Fallback rule only]

⚠️ Limitations

🚧 Next Iterations