CapCheck AI Technical Architecture

Public documentation for the CapCheck AI fact-checking platform. This repo contains architecture decisions, design patterns, and technical deep-dives without proprietary code.

Overview

CapCheck is a fact-checking platform that uses a LangGraph-based multi-agent pipeline to verify claims in real-time. The system combines:

• Claude (Sonnet for reasoning, Haiku for classification/OCR)
• Perplexity Sonar Pro for fact-checking (#1 on SimpleQA benchmark, 0.858 F-score)
• Vision Transformer for AI-generated image detection
• OpenAI Whisper for audio transcription

Architecture Overview

┌─────────────────────────────────────────────────────────────────────────────┐
│                    CapCheck V2 Verification Pipeline                         │
└─────────────────────────────────────────────────────────────────────────────┘

  Request                                                              Response
     │                                                                     ▲
     ▼                                                                     │
┌─────────┐   ┌──────────┐   ┌──────────┐   ┌───────────────┐   ┌──────────┐
│ INTAKE  │──▶│ CLASSIFY │──▶│ STRATEGY │──▶│  SPECIALISTS  │──▶│SYNTHESIS │
│         │   │          │   │          │   │   (parallel)  │   │          │
│ • OCR   │   │ • Satire │   │ • Agent  │   │ ┌───────────┐ │   │ • Verdict│
│ • Audio │   │ • Stakes │   │   select │   │ │web_search │ │   │ • Conf.  │
│ • URL   │   │ • AI rel │   │ • Thresh │   │ │fact_check │ │   │ • Review │
└─────────┘   └────┬─────┘   └──────────┘   │ │ai_detect  │ │   └──────────┘
                   │                        │ │claim_extr │ │
                   │ satire?                │ └───────────┘ │
                   ▼                        └───────────────┘
              ┌──────────┐
              │  OUTPUT  │  ◀─── Short-circuit for satire
              │ (format) │
              └──────────┘

Key Technical Patterns

1. Evidence-Only Synthesis

All verdicts are determined from retrieved evidence only. The LLM is explicitly constrained from using training knowledge to prevent hallucinations.

→ See decisions/001_EVIDENCE_ONLY_SYNTHESIS.md

2. Confidence Calibration

The system detects when LLMs output default confidence values (like 60%) and recalculates based on source quality, quantity, and consensus.

→ See decisions/002_CONFIDENCE_CALIBRATION.md)

3. Model Selection Strategy

Different models for different tasks based on latency, cost, and capability requirements.

→ See decisions/003_MODEL_SELECTION.md

4. Graceful Degradation

Automatic fallback chain when services fail: V2 → V1, Perplexity → Brave Search + Claude.

→ See decisions/04_GRACEFUL_DEGRADATION.md

Documentation

Document	Description
ARCHITECTURE.md	Full system architecture and data flow
LLM_PATTERNS.md	LLM integration patterns and prompts
decisions/	Architecture Decision Records (ADRs)

Tech Stack

Backend: TypeScript, Node.js, Fastify, LangGraph, PostgreSQL, Hasura GraphQL

AI/ML: Claude Sonnet 4, Claude Haiku 3.5, Perplexity Sonar Pro, OpenAI Whisper, Vision Transformer

Frontend: iOS (Swift/SwiftUI), Web (Svelte 5/SvelteKit), Chrome Extension (planned)

Infrastructure: Docker, Fly.io, Cloudflare

Performance

Metric	Value
Cost per verification	$0.025-0.045
Latency (P95)	40-60s full pipeline
Caching	Database-backed with perceptual image hashing

Why This Architecture?

Fact-checking requires accuracy over speed and transparency over black-box verdicts. Key design principles:

1. Evidence-grounded - Never trust LLM training data for factual claims
2. Auditable - Every verdict has traceable source citations
3. Calibrated - Confidence scores reflect actual evidence quality
4. Resilient - Graceful degradation when services fail

Contact

Aaron Kantrowitz
aaronkantrowitz.com | LinkedIn

---

This is a portfolio documentation repo. The CapCheck codebase is private.