🚀 Code Anomaly Engine

Real-time AI-powered vulnerability detection for high-velocity codebases.

---

📌 Overview

Code Anomaly Engine is a polyglot, event-driven microservice architecture designed to intercept GitHub push events in real-time, extract code diffs, and classify them for security vulnerabilities using a fine-tuned CodeBERT transformer model.

It is engineered to handle high-throughput development environments, providing sub-millisecond inference latency and real-time WebSocket updates to a centralized monitoring dashboard.

🏗 Architecture

The system leverages an event-driven design decoupled via Apache Kafka, allowing components to scale independently.

graph TD
    GH[GitHub Webhook] -->|JSON Push Event| IS[Ingestion Service <br/><i>Go</i>]
    IS -->|Parse & Chunk| IS2[Code Snippets]
    IS2 -->|Produce| K[Apache Kafka <br/><i>KRaft Mode</i>]
    
    K -->|Consume| GW[Gateway Service <br/><i>Go</i>]
    GW -->|gRPC Stream| IE[Inference Engine <br/><i>Rust / ONNX</i>]
    IE -->|Predict Vulnerability| GW
    
    GW -->|WebSocket Broadcast| UI[Real-time Dashboard <br/><i>React / TS</i>]

Loading

Components

1. Ingestion Service (Go): Receives GitHub webhook payloads, parses git diffs to extract added/modified code chunks, and publishes structured messages to Kafka.
2. Kafka Event Bus: Decouples ingestion from inference, absorbing sudden traffic spikes during CI/CD surges.
3. Gateway Service (Go): Consumes Kafka messages, acts as a gRPC client to request predictions from the ML backend, and broadcasts results to connected UI clients via WebSockets.
4. Inference Engine (Rust): A high-performance gRPC server utilizing ort (ONNX Runtime) for ultra-fast, sub-50ms ML inference on a CodeBERT sequence classification model.
5. Dashboard (React/TypeScript): Connects to the Gateway via WebSockets to provide a live, real-time feed of code anomalies with syntax-highlighted diffs and latency metrics.
6. ML Pipeline (Python): Downloads the Devign vulnerability dataset, fine-tunes Microsoft's CodeBERT, and exports the computational graph to an ONNX format with dynamic sequence lengths.

🚀 Quick Start

The entire system is containerized. To spin up the Kafka broker, microservices, and dashboard:

docker-compose up --build -d

Services Map

Service	Port	Description
Dashboard	http://localhost:5173	Real-time React frontend
Ingestion	http://localhost:8080/webhook	GitHub webhook receiver
Gateway	http://localhost:8081/ws	WebSocket endpoint for UI
Inference	grpc://localhost:50051	Rust ONNX gRPC Backend
Kafka	localhost:9092	Internal event broker

🧠 Machine Learning

The anomaly detection model is based on Microsoft's CodeBERT fine-tuned on the Devign Dataset (27k vulnerable and clean C/C++ functions).

To recreate the model locally:

make train   # Downloads dataset, tokenizes, and fine-tunes CodeBERT
make export  # Exports the PyTorch model to ONNX with dynamic axes

🛠 Tech Stack Deep Dive

• Go (1.22): Used for IO-bound microservices (Ingestion, Gateway). Leverages goroutines for concurrent webhook processing and Kafka consumption.
• Rust (1.77): Used for CPU-bound ML inference. tonic for async gRPC and ort for highly optimized C++ backend ONNX inference.
• Kafka (KRaft): Used without Zookeeper for simplified infrastructure. Provides fault tolerance and buffering.
• gRPC / Protobuf: Enables strict-schema, low-overhead binary communication between the Go Gateway and Rust Inference Engine.
• React + Recharts: Provides dynamic, real-time metrics (p99 latency) and live UI updates without polling.