VeriFact

Real-time fact verification for online meetings

Live Transcript

Live

Welcome to today's quarterly review meeting.

Our Q3 revenue increased by 23% compared t

This growth was driven primarily by our new product line.

We expect this trend to continue into Q4.

Swipe will appear here when a claim is verified.

Live Transcript

Live

Welcome to today's quarterly review meeting.

Our Q3 revenue increased by 23% compared t

This growth was driven primarily by our new product line.

We expect this trend to continue into Q4.

The Challenge

Meetings are noisy. Facts shouldn't be.

Time lost finding info

9.3 h/week

Cost of inaccurate data

$13M/yr

AI already in use at work

75%

Unverified claims derail decisions and burn hours on manual fact-checking.

Live claim detection with instant source citations; right inside your call.

VeriFact auto-joins your Zoom calls to detect and verify claims in real time. No extensive setup required.

A seamless pipeline from speech to verified truth

Speech Input

Zoom Bot + STT (Whisper)

Claim Detection

BERT + FEVER Dataset

RAG Retrieval

Embeddings + Graph (Neo4j)

Verification

Classifier: Support/Refute/NEI

Display

Zoom App Pane + Web UI

Zoom bot joins meetings, captures separate audio per participant, and identifies speakers for clear transcripts.
Technical Details
Participants can invite the bot to scheduled or ongoing meetings. Separate audio streams enable speaker attribution; transcribed via a robust STT model (e.g., Whisper).
Fine-tuned transformer identifies factual claims in real-time text.
Technical Details
BERT (~110M params, ~440MB) fine-tuned on FEVER (185k+ claims). Target hardware ≥12GB VRAM for efficient training and inference with small batches.
Hybrid retrieval fetches evidence using semantic vectors and graph relations.
Technical Details
Retriever combines embedding search with graph-based links (Neo4j). Containerized services + caching for sub‑second retrieval; data encrypted to align with GDPR.
Verifies each detected claim against retrieved evidence.
Technical Details
Outputs Supported, Refuted, or Not Enough Info. Success measured by precision/recall; end‑to‑end latency targeted within a few seconds per sentence.
Live pane shows transcript, detections, and verifications; users can manage docs and ask questions.
Technical Details
Zoom side pane provides real-time feedback. App page enables uploads/management of documents. Q&A via trigger phrase "Hey Agent" consults the document pool.