Our Methodology

Zero Data Retention Architecture

NewsAssist AI is founded on Client-Side Confidentiality. No user content (uploaded files, transcripts, or analysis results) is ever stored on our backend servers. All processing is conducted in-memory and drafts are stored locally on your device (using client-side IndexedDB).

Data Immutability (The Audit Log)

The Audit Log (metadata only: who, what, when, status) is stored in Firestore to provide non-repudiable proof that your confidential content was purged post-analysis.

Model Training Exclusion

We do not, and will never, use any user-uploaded proprietary data to train or refine our underlying Large Language Models (LLMs).

A. The Multilingual Engine (36+ Languages)

We utilize a two-stage API pipeline for cross-language content analysis, ensuring accurate source identification (including low-resource languages like Nigerian Pidgin) and delivering final output in the user's selected language.

B. Verification and Extraction Methodologies

• Entity & Precedent Finder (Legal): Uses **Named Entity Recognition (NER)** models specifically tuned to identify legal entities and structured patterns.
• Citation Integrity Check (Academic): Uses **Structural Pattern Matching** to confirm the relationship between sources cited in the bibliography and the claims in the document body.
• Source Verification (Journalism): Uses a **String Matching Algorithm** to directly compare quotes and data points in the draft story against the original secure source file loaded locally.

We employ techniques like **token-level filtering** and model temperature control to reduce the likelihood of biased or speculative outputs, ensuring the AI maintains a neutral, professional, and objective tone.

Transparency of Limitations

The AI structures and accelerates the process; the user is always responsible for the final output.