How satwik77-dev/uf-eee6778-fall25-termproject works
Project Type & Purpose
• ClaimVerify is a full-stack AI-powered fact-verification system.
• It ingests user claims via a Streamlit UI, runs an inference pipeline combining offline fact-check databases (PolitiFact, Snopes) with real-time Google search fallbacks, and returns verdicts plus interpretability insights.
Top-Level Components
a. UI Layer
• streamlit_app.py
• Theme assets (dark_theme.css, light_theme.css)
b. Backend Inference Pipeline
• src/preprocess.py: Text cleaning & harmonization
• src/embeddings.py: Sentence-BERT embedding generation
• src/retrieval.py: FAISS index search + Google API fallback
• src/model.py: Classifier loading & inference (RoBERTa fine-tuned)
• src/explainability.py: SHAP interpretability
• src/inference_pipeline.py: Orchestrator that ties modules together
c. Data Storage
• data/raw/: Original PolitiFact & Snopes datasets
• data/processed/: Cleaned/merged CSV, .npy/.pkl embeddings, FAISS index & metadata
d. Model Artifacts
• models/classifier/roberta_finetuned/: Model checkpoints, tokenizer, label mapping, temperature scaling
e. External Services
• Google Custom Search API (for live fact-check articles)
f. Supporting Notebooks & Reports
• notebooks/: EDA, offline retrieval, training
• DeliverableReports/ & results/: Documentation & plots
Relationships & Data Flow
• User enters claim → UI Layer
→ UI calls Inference Pipeline API (local function call in Streamlit)
→ Preprocess cleans claim text
→ Embeddings module encodes to semantic vectors
→ Retrieval searches FAISS index (offline_db).
• If no match above threshold → trigger Google API search
→ Classification model predicts credibility label
→ Explainability module generates SHAP visuals
→ Pipeline returns: similar claims, verdict, confidence, interpretability plots
→ UI renders results
Architectural Patterns & Design Principles
• Layered Architecture: Presentation (Streamlit), Application (inference modules), Data (processed datasets, indices)
• Modular Code: Each src/*.py implements one concern
• Reusable Artifacts: Precomputed embeddings & FAISS index for performance
• Failover Strategy: Offline-first with API fallback for resilience
• Separation of Concerns: Notebooks for exploration; src/ for production logic; ui/ for front-end
Key Technologies & Frameworks
• Python 3.9+
• Streamlit (UI)
• Sentence-BERT (embeddings)
• FAISS (vector search)
• HuggingFace Transformers (RoBERTa)
• SHAP (interpretability)
• Pandas, NumPy (data processing)
• Google Custom Search API