An intelligent onboarding accelerator that transforms dense repositories into interactive guides, drastically reducing engineer ramp-up time.
Live Demo · GitHub · Features · Tech Stack · Architecture
- Dynamic Executive Summaries — Instantly unpacks new codebases. Automatically detects tech stacks, identifies primary application entry points, and traces data flows.
- Dependency Inference — Synthesizes module and directory-level summaries to give new engineers a bird's-eye view.
- Topological Learning Paths — Uses Abstract Syntax Tree (AST) import dependency graphs and Tarjan's algorithm to compute foundational modules vs. entry points.
- Where To Begin — Provides a step-by-step checklist annotated by an LLM ("Read
database.pyfirst because it sets up the DB..."), entirely removing the guesswork.
- Domain Dictionary — Extracts functions, classes, and complex variable structures and runs them against AI to generate plain-English explanations.
- Searchable Definitions — New hires can immediately search what a niche domain term or custom service class means, without hunting down the original author.
- Git Intelligence Integration — Leverages deep
git logand GitHub Pull Request analysis to attribute primary owners and map out contribution percentages. - Knowledge Silo Flags — Automatically identifies bus-factors of 1 (files only touched by a single engineer) and traces domain expertise.
- Traffic-Light Prioritization — Tags core plumbing and infrastructure components (🔴 Critical, 🟡 Caution, 🟢 Safe to Explore) using heuristic algorithms that measure code fan-in, historical test presence, and update churn rates.
- Multi-Source Context Search — Vector embeddings generated off AST symbols, semantic code, pull request discussions, and commit messages.
- The "Why" Beyond the "What" — Empowers queries like "Why did we switch to gRPC in the messaging service?" by referencing the original pull request context rather than just blind code reading.
- Interactive Visualization — Render complex dependencies natively in the browser via WebGL and
react-force-graph-3d. Nodes are colored by criticality and sized relative to their architectural weight.
| Layer | Technology |
|---|---|
| Backend Framework | FastAPI (Python 3.12, Async) |
| Frontend UI | Next.js 15 (App Router, Tailwind CSS 4) |
| Parsing Engine | Tree-Sitter |
| Database | PostgreSQL + SQLAlchemy 2.0 |
| Vector Search | pgvector + OpenAI Embeddings |
| Task Distributed Queue | Celery + Redis |
| Visual Graphing | react-force-graph-3d |
| AI LLM | OpenAI API |
- Python 3.11+
- Node.js 18+
- Docker & Docker Compose
- OpenAI API Key & GitHub OAuth Credentials
# 1. Spin up the Postgres (with pgvector) and Redis instances
docker-compose up -d
# 2. Launch the FastAPI backend
cd server
uv sync
uv run uvicorn app.main:app --reload
# 3. Launch the Celery Worker pipeline
uv run celery -A app.core.celery worker --loglevel=info
# 4. Boot the Next.js Client
cd client
npm install
npm run devgraph TD
Client[Next.js Frontend] -->|REST / WebSockets| API[FastAPI Server]
API -->|Queue Repository| Redis[Redis Broker / PubSub]
Worker[Celery Pipeline] -->|Clone & Analyze Git| GitHub[GitHub API]
Worker -->|AST Extraction| Scanner[Tree-sitter Engine]
Scanner --> Graph[PostgreSQL]
Worker -->|Chunking & Summaries| LLM[OpenAI API]
LLM --> Vector[(pgvector Storage)]
Worker -->|Streaming Telemetry| Redis
API -->|Subscribe to Status| Redis
Built with ❤️ by Tejas