← All Projects
07 · Project

Anviksha

Multi-agent AI code review system where specialist agents analyze security, performance, and architecture in parallel, with blind evaluation and conflict-aware reasoning.

LangGraphFastAPIGroqPostgreSQLNext.jsAuthShield
↗ GitHub▶ Live Demo
Overview

Anviksha is a multi-agent AI code review system designed to separate security, performance, and architectural reasoning into independent specialist agents instead of relying on a single generic LLM response.

An orchestrator first analyzes the codebase and prepares tailored context for each agent before the specialists run in parallel. Their findings are then passed through a blind evaluator that scores relevance, accuracy, and actionability without knowing which agent produced the result — reducing evaluation bias and improving signal quality.

A conflict detection layer identifies contradictions between findings, while an on-demand Judge Agent resolves disputes only when requested, avoiding unnecessary latency in the primary pipeline.

The system is built on LangGraph with SSE-based progressive streaming, allowing results to appear in real time as agents complete. Running on Groq LPUs, the full multi-agent pipeline completes in under 5 seconds while maintaining interactive UX.

3Specialist AI agents
<5sFull pipeline latency
SSEReal-time result streaming
BlindEvaluator scoring system
Tech Stack

TOOLS &
TECHNOLOGIES.

TECH.01
LangGraph
TECH.02
FastAPI
TECH.03
Groq
TECH.04
PostgreSQL
TECH.05
Next.js
Key Highlights

WHAT MAKES IT
STAND OUT.

3Parallel specialist agents
<5sEnd-to-end analysis latency
BlindEvaluator scoring system
SSEProgressive result streaming
2Modes: Try Demo and authenticated dashboard
JudgeOn-demand conflict resolution
Challenges & Solutions

WHAT WAS HARD.
HOW I SOLVED IT.

Critical Challenge

Single-model code review produces generic feedback where security, performance, and architecture concerns blend together.

Engineered Solution

Introduced specialist agents with isolated context frames generated by an orchestrator, preventing domain bleed and keeping each agent focused on its own concern.

Critical Challenge

Evaluator bias can distort scoring when the evaluator knows which agent generated a finding.

Engineered Solution

Implemented blind evaluation by stripping agent identity before scoring relevance, accuracy, and actionability across all findings.

Critical Challenge

Inline multi-agent debate significantly increased latency and blocked progressive result streaming.

Engineered Solution

Moved the Judge Agent outside the main LangGraph pipeline and triggered it on-demand only when users request conflict arbitration.

Core Architecture

HOW IT'S
ALL WIRED UP.