AI-Driven Pull Request Security Review Platform

Project Overview

This project is an enterprise-grade, automated Application Security (AppSec) review platform designed to continuously monitor source control activity and apply AI-powered analysis to every code change before it is merged. The core problem it solves is the scalability gap between security engineering capacity and engineering throughput: as development teams grow, manual code review for security vulnerabilities becomes a bottleneck. This platform eliminates that bottleneck by running autonomous, AI-driven security analysis on pull requests across multiple repositories and organizations, 24/7, without human intervention.

The platform integrates directly with a source control management API to discover new and updated pull requests in near real-time. Each discovered PR is queued and processed through a multi-phase AI pipeline that analyzes changed files for vulnerabilities — including injection flaws, authentication bypasses, insecure deserialization, and secrets exposure — before synthesizing a structured verdict. The resulting security review, including file-level findings and a final risk classification, is persisted and surfaced through a web-based management dashboard.

Beyond passive reporting, the platform enforces active gatekeeping: it can block PR merges automatically when a high-risk verdict is issued by posting a blocking status check. This transforms security review from an advisory process into an enforceable compliance control, delivering measurable reduction in the risk of vulnerable code reaching production.

Key Features

• Continuous PR Discovery Engine: A resilient background service continuously polls the source control API for open pull requests across all configured organizations and repositories, applying a configurable look-back window and deduplication logic to avoid redundant processing.

• Multi-Phase AI Security Analysis Pipeline: Changed files are processed through a token-aware chunking algorithm and submitted to a large language model in two phases — per-file analysis and a final cross-file synthesis — producing structured JSON verdict objects with detailed reasoning and remediation recommendations.

• Automated Merge Gatekeeping (Policy as Code) Security alerts are useless if developers can bypass them to hit a release deadline. This platform eliminates that loophole by translating static compliance rules into executable Policy as Code.
  
  By injecting real-time status checks directly into your VCS, the system automatically blocks merges on flagged PRs. It enforces rigid workflow SLAs and shuts the gate instantly the moment a vulnerability is detected. The result? A pristine main branch and a security team that no longer has to play manual babysitter.
  
  But handing over the keys to an automated gatekeeper introduces a new, critical risk. If the pipeline is making hard merge decisions, the AI analyzer itself must be bulletproof against malicious manipulation.

• Prompt Injection & Output Sanitization: AI prompt inputs (file names, patch content, PR titles) are sanitized before construction, and all AI-generated outputs rendered in the web UI are HTML-encoded by default. GitHub comment content undergoes additional Markdown neutralization to prevent downstream injection.

• Admin Dashboard with Runtime Controls: A web-based management interface allows operators to monitor active reviews, inspect per-file analysis details, view platform metrics, and dynamically toggle settings — such as the active AI model and gatekeeping mode — without restarting the service.

• Disaster Recovery CLI Tool: A companion command-line utility enables operators to replay or force-trigger reviews for specific repositories and PRs, supporting incident recovery and audit scenarios without requiring direct database access.

Tech Stack

Backend

• Runtime: .NET 8 (ASP.NET Core MVC), C# 12

• AI Integration: Google Generative AI SDK (Google.GenAI) for structured content generation using configurable large language models

• Source Control Integration: Octokit and Octokit.Reactive for GitHub REST API interaction; GitHubJwt for Asymmetric Key-based App Authentication

• Background Processing: IHostedService-based orchestration with a producer/consumer Channel<T> queue, watchdog-wrapped loops with exponential backoff restart, and SemaphoreSlim-bounded concurrency (max 5 concurrent AI jobs)

• Authentication: OpenID Connect (Microsoft.AspNetCore.Authentication.OpenIdConnect) with data protection backed by Entity Framework Core

• ORM & Database: Entity Framework Core 9 with Npgsql provider targeting PostgreSQL

Frontend

• Rendering Engine: Razor Views (ASP.NET Core MVC server-side rendering)

• Security: Default Razor HTML encoding on all AI-generated and user-controlled output; URL scheme validation on external links; rel="noopener noreferrer" on all outbound anchors

DevOps & Infrastructure

• Containerization: Docker with Linux target OS; multi-service local environment via Docker Compose

• Orchestration: Kubernetes deployment via Helm Chart, with configmap-backed application settings and secret volume mounts for private key material

• Secrets Management: Zero-hardcoded-secret policy; all credentials injected via environment variables or Kubernetes secrets at runtime

Conclusion

The platform demonstrates a mature architectural approach to embedding security automation into the software development lifecycle. The deliberate decoupling of PR discovery from review execution — via a bounded in-memory channel — ensures the system remains responsive and backpressure-safe under high load. The multi-phase AI pipeline with token-aware chunking addresses the practical challenge of reviewing large changesets that exceed single-prompt context windows, while the watchdog-wrapped background loops provide resilience against transient external API failures.

The consistent application of zero-trust principles — prompt sanitization, output encoding, secrets injection, and log redaction — ensures the platform itself does not become an attack surface despite processing untrusted, externally-sourced code content. The result is a production-hardened tool that raises the security baseline across an engineering organization without introducing friction into the development workflow.