The Multi-Agent AI Code Auditor & Mentoring System
AlgoShield is an enterprise-grade AI developer tool that helps students and engineers debug C++ code by simulating a Panel of Judges. Rather than simply providing solutions, AlgoShield parses entire workspaces, evaluates algorithmic complexity, and provides pedagogical hints—all while enforcing strict pedagogical guardrails.
- Multi-Agent Consensus Panel: Instead of a single LLM, your code is processed by a 4-agent workflow (Complexity Judge, Systems Judge, Aggregator, and Critic) ensuring high-quality, safe, and educational feedback.
- Workspace-Aware RAG: Unlike standard chatbots, AlgoShield tokenizes entire C++ workspaces, building a structural dependency blueprint to understand how your headers, utils, and main files interact.
- Self-Healing Architecture: Implements a feedback loop where an AI Critic reviews the generated hint. If the hint fails quality standards, the graph auto-corrects and regenerates the response.
- Production-Ready CI/CD: Fully containerized with Docker Compose and integrated with GitHub Actions for automated quality assurance on every push.
The system utilizes a directed acyclic graph (DAG) for agent orchestration:
- Workspace Parser: Tokenizes and builds an AST-like dependency blueprint.
- Judge Panel: Independent agents run parallel analysis on performance and memory constraints.
- Aggregator & Critic: Synthesizes findings into a pedagogical hint while performing security audits to prevent prompt injection and "answer-spoiling."
- AI Orchestration: LangGraph
- LLM Engine: Groq (Llama 3.1)
- Vector Database: Qdrant
- Backend: FastAPI
- Frontend: React + Tailwind CSS
- Infrastructure: Docker, Docker Compose, GitHub Actions