Self-Initiated R&D · End-to-End System Design

CodeIntel

AI-native development infrastructure — automated PR reviews, semantic code search, and documentation generation. Built from scratch with Go, TypeScript, and Claude AI.

2 Hetzner Servers
5 Languages Parsed
Full-Stack Go + TS

Engineering Philosophy

Built with Quality at Every Layer

From automated code reviews to real-time observability — every component is designed to catch issues early, ensure reliability, and maintain confidence in production.

Automated Quality Gates

Every PR triggers automated analysis — Claude reviews diffs with full call graph context, catching bugs, security issues, and anti-patterns before merge.

Deep Debugging Tools

Real-time log streaming, database introspection, vector store explorer, and connection pool monitoring — all accessible through a custom admin dashboard.

CI/CD Pipeline Design

Webhook-driven async pipelines with Asynq queues, idempotent job processing, exponential backoff retry, and Check Runs status reporting.

Reliability Engineering

Systemd service management with boot ordering, private network isolation, auto-TLS via Caddy, and bare repo caching with per-repo mutexes.

Observability First

WebSocket-powered real-time dashboards for queue metrics, server health, job processing status, and system resource monitoring.

Confidence-Based Resolution

Hybrid static + TypeScript API + heuristic pipeline with 100%/95%/70% confidence scoring — ensures call graph accuracy at scale.

Core Features

What It Does

A self-hosted system for a team of 20+ developers across 100+ repositories — automating code quality at every stage.

Automated PR Reviews

GitHub App receives webhooks, queues review jobs, and posts inline comments via Claude Agent SDK. Structured output with severity, category, and actionable suggestions.

Semantic Code Search

Voyage Code 2 embeddings indexed in Qdrant. Multi-factor relevance scoring combines call graph distance, semantic similarity, file proximity, and historical significance.

Deep Code Understanding

Tree-sitter parsing across Go, TypeScript, JavaScript, PHP, and Python. Call graph construction with confidence-based resolution and cross-repo dependency tracking.

AI-Powered Documentation

Claude generates MDX documentation, creates PRs via GitHub App, and detects stale docs. Static site aggregates docs from all repositories at build time.

Infrastructure

Two-Server Architecture

Data and compute isolated for reliability and security — connected via private network on Hetzner Cloud.

Data Server

CPX32 · 4 vCPU · 8GB RAM

Go API

REST + WebSocket

Asynq Worker

Background jobs

PostgreSQL 16

Relational data

Qdrant

Vector search

Redis

Queue & cache

Caddy

HTTPS + reverse proxy

Compute Server

CPX22 · 2 vCPU · 4GB RAM

Review Worker

PR analysis

Doc Worker

Doc generation

Claude Agent SDK

AI integration

GitHub App

Webhook receiver

Private network (10.0.0.0/16)

Automation Pipeline

PR Review Flow

End-to-end automated review — from git push to inline GitHub comments in under 2 minutes.

Push / PR

Developer opens PR

Webhook

GitHub App notified

Queue

Asynq job created

Worker

Fetches PR context

Claude Review

AI analyzes diff + graph

PR Comments

Inline feedback posted

Admin Dashboard

Full Observability

Custom-built admin dashboard with real-time monitoring, database introspection, and system management — built with Next.js 16, React 19, and WebSocket.

System Overview
1 / 6

System Overview

Real-time monitoring of repositories, job queues, server health, and processing activity

React 19 · shadcn/ui · TanStack Table · Recharts · WebSocket

Under the Hood

Engineering Highlights

Tree-sitter Parsing

5 languages with AST-level extraction of functions, classes, imports, and call sites. Thread-safe parser management.

Hybrid Resolution Pipeline

Static (100%) → TypeScript API (95%) → Heuristic (70%) for call graph accuracy at scale.

PageRank Repo Maps

Importance-weighted repository overviews using call graph centrality for smarter context selection.

Token Budget Management

32K token context assembly with priority-based selection from semantic search results.

Webhook Idempotency

Asynq TaskID deduplication prevents duplicate reviews. Exponential backoff with max 2 retries.

Bare Repo Cache

Git bare repos + worktrees for thread-safe concurrent access with per-repo mutexes.

Technologies

Tech Stack

GoTypeScriptPostgreSQLQdrantRedisTree-sitterClaude AIVoyage AINext.jsCaddyHetzner CloudGitHub App APIAsynqDockersystemd