Lumira
Voice-activated AI guide for exhibitions with context-locking architecture
Llama 3.2
LLM
mxbai-embed-large
Embeddings
ChromaDB
Vector DB
Zero
Hallucination Rate
The Problem
At project expos and viva presentations, students panic when asked unexpected questions about their own projects. Documentation is scattered, and there's no fast way to query specific technical details from your own project report.
The Solution
LUMIRA is a context-aware AI-powered project assistant built for expos and exhibitions. Give it a PDF, it studies the content, vectorizes it, and answers questions like a pro — with zero hallucinations. Built on a RAG (Retrieval-Augmented Generation) pipeline using LangChain, Llama 3.2, and ChromaDB.
System Architecture
PDF → Vector → LLM pipeline. PyPDFLoader extracts and sections the PDF (Executive Summary, Methodology, etc.). mxbai-embed-large creates vector embeddings of each section. ChromaDB stores the vectorized content. At query time, relevant context is retrieved from the vector store and fed to Llama 3.2 via LangChain, which generates responses based strictly on the project PDF — no hallucinations.
Key Features
PDF ingestion with automatic section extraction
Vector embedding using mxbai-embed-large model
ChromaDB vector store for fast similarity search
Context-locked responses — answers only from your PDF data
Zero-hallucination Llama 3.2 inference via Ollama
Lightweight, local, and fully private — no cloud dependencies
My Role
Sole developer — designed the RAG pipeline, built the vectorization engine, integrated Llama 3.2 via Ollama, and developed both the Python backend and React frontend.