CapyNodes

CapyNodes is a technical platform designed for software engineers to practice system design through interactive diagramming and real-time collaboration. It features an advanced “Judge” system that provides instant, detailed feedback on architectural decisions.

---

Demo: System in Action

Demo: Features

---

Backend Architecture

The CapyNodes backend is built on a robust hybrid architecture, combining the stability of Django with the real-time capabilities of Django Channels.

• Core Stack: Python 3.x, Django, PostgreSQL (via NeonDB).
• Real-time Engine: Django Channels with a Redis channel layer for sub-100ms message broadcasting.

---

AI Evaluation Engine

The “Judge” system uses a multi-stage pipeline to provide consistent and high-fidelity scoring of system design diagrams.

1. Stage 1: Rule-Based Validation: Detects basic structural issues and detected anti-patterns like single points of failure.
2. Stage 2: LLM Chain-of-Thought: Leverages Qwen 32B (via Groq) for deep architectural reasoning, identifying strengths and weaknesses across dimensions like scalability and performance.
3. Stage 3: Score Aggregation: Normalizes heuristics and LLM feedback into a final, multi-dimensional score.

---

Real-time Collaboration

The collaboration system allows multiple users to co-edit diagrams with seamless synchronization and live cursor tracking.

• WebSocket Protocol: Custom consumers handle state_update for persistent changes and cursor_update for high-frequency tracking.
• Performance Optimization: Cursor updates are broadcasted directly without database persistence to minimize latency. Database synchronization for diagram states is handled asynchronously using database_sync_to_async.
• Conflict Resolution: Sender-side filtering and immediate broadcasts ensure a smooth, feedback-loop-free experience for all participants.

---

Observability & Quality

To maintain the accuracy of the AI judge, CapyNodes includes a dedicated observability infrastructure.

• Streamlit Dashboard: Monitor quality metrics, LLM token costs, and P95 latency in real-time.
• Golden Tests: A curated “Golden Set” of diagrams is used for regression testing whenever prompts or models are updated.
• Hallucination Monitoring: Secondary models (Gemini 1.5 Pro) are used offline to validate the primary judge’s scoring consistency.