‹ All Projects
Project Thread🏆 Inventx 2025 - Gold

AI Evaluator — End-to-end judging platform with multi-LLM auto-review pipeline

🐙 View GitHub

Overview

Event organizers upload rubrics; participants submit proposals (PDF / image). LLM ensemble filters, scores and ranks submissions.

Tech Stack

ReactNode.jsTesseract-OCROpenAI, Gemini, Claude APIsMongoDB

Features

  • Multi-tenant event setup
  • OCR preprocessing of images
  • Rule-based eligibility filter
  • LLM ensemble summary & feedback
  • Auto-generated scorecards & rankings

Key Decisions

01

Implemented a "Conflict of Interest" detector that flags submissions with similar semantic patterns to the judging rubrics to prevent prompt injection.

02

Architected an asynchronous polling system in Node.js to handle long-running ensemble calls across three different model providers simultaneously.

Learnings

→Discovered that mathematically averaging scores across three model architectures (GPT, Claude, Gemini) reduces outlier bias by 25% compared to single-model judging.
→Gained experience in normalizing disparate JSON schemas from different AI providers into a unified internal scorecard format.
EmailGitHubLinkedInResume