case://study-003
RAG / NLPProject 003
IdeaLens AI
IdeaLens AI helps screen hackathon submissions at scale by combining template-aware PDF chunking, embedding-based retrieval, and strict JSON scoring grounded in rubric evidence.
100 pts
Rubric
5
Sections
384-d
Embeddings
RAG / NLP
Category
The Problem
What Wasn't Working
Judging large numbers of proposals manually is slow and inconsistent, especially when evaluators need evidence-backed scoring across multiple rubric dimensions.
The Solution
How I Fixed It
I designed a rubric-aware pipeline that chunks each PDF into five sections, stores 384-dimensional embeddings in ChromaDB, retrieves supporting evidence per criterion, and enforces strict JSON scoring with validation and invalid-submission checks.
Stack
Technologies Used
Python
LLaMA-3
ChromaDB
Sentence Transformers
NLP
RAG
Results
Key Outcomes
✓
Template-aware chunking across five rubric sections✓
Evidence-grounded scoring with vector retrieval✓
Near-duplicate detection for submission quality control✓
Strict JSON output on a 100-point rubric