VeriAlign

● Active

A reverse proxy that intercepts LLM API calls and augments responses with verification metadata.

Python Reverse Proxy LLM APIs Claim Extraction Source Grounding Contradiction Detection
GitHub →

Overview

VeriAlign sits between any application and any OpenAI-compatible provider, intercepting LLM API calls and augmenting responses with verification metadata. Every claim the model makes gets traced, grounded, and scored.

What it does:

  • Claim Extraction — Parses model output into discrete, testable claims
  • Source Grounding — Maps each claim to supporting evidence or flags unsupported assertions
  • Contradiction Detection — Identifies internal inconsistencies within a single response
  • Confidence Scoring — Assigns reliability scores to each claim based on grounding evidence
  • Checklist Verification — Validates responses against predefined compliance checklists

How It Works

┌─────────────┐     ┌──────────────┐     ┌─────────────────┐
│  Your App   │────▶│  VeriAlign   │────▶│  LLM Provider   │
│             │◀────│  (Proxy)     │◀────│  (OpenAI, etc.) │
└─────────────┘     └──────────────┘     └─────────────────┘
                           │
                    ┌──────▼──────┐
                    │  Verified   │
                    │  Response   │
                    │  + Metadata │
                    └─────────────┘

Architecture

The proxy intercepts HTTP requests to any OpenAI-compatible endpoint, forwards them to the actual provider, then post-processes the response:

  1. Intercept — Catches the streaming or non-streaming response
  2. Extract — Parses claims from the model output using NLP pipelines
  3. Ground — Cross-references claims against provided context or knowledge base
  4. Score — Computes confidence based on grounding strength and contradiction analysis
  5. Augment — Attaches verification metadata to the response before returning it

Getting Started

# Clone the repository
git clone https://github.com/Anandhasasidharan/verialign.git
cd verialign

# Install dependencies
pip install -r requirements.txt

# Run the proxy
python -m verialign --port 8080 --upstream https://api.openai.com

Use Cases

  • AI Safety — Verify model outputs before acting on them
  • Compliance — Ensure responses meet regulatory checklists
  • Red Teaming — Detect contradictions and unsupported claims in adversarial contexts
  • Blue Team Ops — Ground security recommendations in real evidence