VeriClaim

Inspiration

In an era flooded with synthetic media, algorithmic content generation, and sophisticated misinformation, distinguishing verifiable facts from plausible falsehoods has become increasingly difficult. Traditional manual fact-checking cannot keep pace with the sheer velocity of information, while generic AI models frequently suffer from "hallucination"—presenting inaccurate claims with unearned confidence.

We were inspired to build VeriClaim to bridge this trust deficit. We envisioned a transparent, automated verification engine that doesn't just pass binary judgment, but actively deconstructs complex narratives, cross-examines authoritative primary sources, and delivers explainable, mathematically grounded credibility assessments that anyone can audit and trust.

What it does

VeriClaim is an AI-powered verification engine engineered to deconstruct complex factual claims, retrieve multi-source authoritative evidence, and deliver explainable credibility assessments in real time.

When a user submits text, documents, or research claims, VeriClaim:

• Deconstructs Atomic Claims: Parses unstructured narrative into discrete, testable propositions.
• Hybrid Evidence Retrieval: Queries trusted knowledge graphs, verified web indexes, and primary data sources via dense vector search and sparse keyword retrieval.
• Computes Explainable Credibility Scores: Formally evaluates semantic entailment, source credibility, and temporal freshness using a weighted scoring model: $$\text{Credibility}(C) = \sum_{i=1}^{n} w_i \cdot \sigma\left(\text{Entailment}(c, e_i) \times \text{Trust}(e_i)\right)$$
• Generates Actionable Reports: Delivers an interactive breakdown of verified points, contradictory evidence, and missing context with direct citation links.

How we built it

• Core Pipeline & NLP Engine: Orchestrated using Python, LangChain, and LlamaIndex with cutting-edge Large Language Models (LLMs) executing strict zero-shot decomposition and reasoning chains.
• Hybrid Retrieval System: Implemented a dual-layer retrieval pipeline pairing dense vector databases (Qdrant / Pinecone) with sparse retrieval (BM25) and live search APIs.
• Backend Architecture: Powered by FastAPI to support asynchronous streaming, concurrent evidence fetching, and deterministic caching.
• Frontend & Visualization: Built with Next.js, React, TypeScript, and Tailwind CSS, featuring an intuitive dashboard with interactive claim-evidence graphs and confidence indicators.

Challenges we ran into

• Contextual Ambiguity & Nuance: Differentiating between strict factual errors, subjective statements, and partially true assertions. We addressed this by deploying a nuanced multi-class verdict taxonomy (Supported, Refuted, Partially True, Unverifiable).
• Latency vs. Accuracy Bottleneck: Exhaustive cross-source evidence verification is compute-heavy. We optimized performance through parallelized async retrieval pipelines and streaming token outputs.
• Source Trust Calibration: Resolving conflicting claims across sources with differing domain authorities by engineering a dynamic, weighted trust heuristic.

Accomplishments that we're proud of

• Zero-Hallucination Framework: Enforced strict citation grounding where every assertion is mapped to primary evidence without autonomous fabrication.
• High-Throughput Parsing: Achieved sub-second claim decomposition across long-form documents.
• Transparent Explainability: Successfully bridged complex statistical confidence metrics with human-interpretable visual badges and reasoning steps.

What we learned

• Hybrid Search Superiority: Combining dense embeddings with exact keyword/entity matching is essential for high-precision fact retrieval (dates, numbers, names).
• Structured Reasoning Chains: Enforcing step-by-step intermediate verification states significantly reduces reasoning drift compared to single-pass prompting.
• User Trust Over Black-Box Numbers: Transparent, verifiable source citations foster greater user adoption than arbitrary percentage scores.

What's next for VeriClaim

• Browser Extension: One-click highlighted text verification directly within social platforms, research portals, and news outlets.
• Multimodal Claim Extraction: Expanding ingestion support to video transcripts, podcast audio feeds, and annotated infographic charts.
• Decentralized Audit Trail: Integrating verifiable credentials and cryptographic attestation proofs to ensure tamper-proof verification history.

Built With

• acdc
• credential
• deepfakes
• gemini
• genai
• generative-ai
• gleif/vlei
• golang
• keri
• pqc
• react
• trust-framework