Axon

Axon is a Truth Maintenance System that continuously verifies company knowledge against reality, detects drift, and automatically proposes fixes before misinformation spreads

Comment

Inspiration

Every engineering team eventually faces the same problem: documentation slowly drifts away from reality.

Configuration changes, APIs evolve, authentication mechanisms change, and infrastructure is modernizedbut documentation is rarely updated at the same pace. Engineers lose trust in internal knowledge and spend valuable time figuring out what is actually true.

We built Axon to solve this problem.

Rather than acting as another documentation bot, Axon is designed as an AI-powered Truth Maintenance System that continuously verifies organizational knowledge against reality and surfaces evidence-backed contradictions before they become costly mistakes.

What it does

Axon connects to a GitHub repository and automatically:

  • Extracts documentation and source code
  • Builds structured knowledge from both
  • Verifies documentation claims against the implementation
  • Detects documentation drift
  • Explains contradictions with supporting evidence
  • Prioritizes findings by severity

Instead of simply searching text, Axon reasons over documentation claims and validates them against the actual implementation.

For our demo, Axon successfully identifies realistic contradictions including:

  • API Keys vs JWT Authentication
  • Disabled feature flags vs enabled implementation
  • SMTP vs SendGrid
  • Additional documentation drift across configuration and infrastructure

Our long-term vision extends beyond GitHub to organizational knowledge sources such as Notion, Slack, Jira and Confluence.

How we built it

Axon was built using:

  • GPT-5.6 and Codex for implementation, debugging, rapid iteration, and development acceleration
  • OpenAI Embeddings for semantic retrieval
  • FastAPI backend
  • Next.js frontend
  • PostgreSQL with pgvector
  • SQLAlchemy + Alembic
  • Docker & Docker Compose
  • AWS EC2 deployment
  • GitHub API integration
  • Background workers for repository analysis

The pipeline:

Repository → Documentation Extraction → Claim Extraction → Semantic Verification → Evidence Generation → Truth Feed

Challenges we ran into

The biggest challenge wasn't building the UI—it was making verification reliable.

We had to solve problems around:

  • Mapping documentation claims to implementation
  • Reducing false positives
  • Reliable repository ingestion
  • Background processing
  • GitHub API limitations
  • Production deployment on AWS

Designing a workflow that produces explainable evidence instead of opaque AI responses was one of the most challenging parts of the project.

Accomplishments that we're proud of

  • Built a complete end-to-end documentation verification pipeline
  • Successfully detect documentation drift using AI
  • Evidence-backed findings instead of black-box outputs
  • Modern production-ready architecture
  • Clean developer experience
  • Successfully deployed on AWS EC2
  • Created a realistic repository to evaluate documentation drift

What we learned

Building Axon taught us that the hardest problem isn't generating code—it's maintaining trust in organizational knowledge.

We also learned how important retrieval quality, semantic verification, and explainability are when building AI systems that developers rely on.

Using GPT-5.6 and Codex dramatically accelerated development, allowing us to iterate quickly on architecture, debugging, deployment, and implementation while keeping our focus on solving the core problem.

What's next for Axon

This submission demonstrates GitHub as Axon's first knowledge source.

Our vision is much larger.

Future versions will connect GitHub, Notion, Slack, Jira, Confluence, Google Docs, and internal documentation into a single continuously verified organizational knowledge graph.

Beyond detecting documentation drift, Axon will become a living company knowledge system capable of answering questions, identifying contradictions across multiple sources, and recommending evidence-backed remediation while keeping humans in control of final decisions.

Built With

Share this project:

Updates