NEXORA: Your Private, Local-First AI Assistant Powered by Open Source

ABHINAV — Wed, 28 Oct 2026 20:23:31 +0000

I built NEXORA an "On-Device AI Workspace", a fully offline, private local-first AI assistant and document intelligence workspace created for the Hacktoberfest Open-Source AI Challenge.

As a mechanical engineering student at NIT Trichy, I wanted a tool that could instantly parse dense research papers, textbook chapters, and codebases without sending any data over the internet or risking privacy leaks. NEXORA combines local model inference through Ollama, semantic vector document search using PyMuPDF and NumPy embeddings, mathematical tool execution, and global capital weather forecasting into a single lightning-fast Streamlit dashboard.

Demo
Here is a video link of NEXORA in action locally:

Functioning snapshots:

Partner Technologies

For this project, I integrated several open-source technologies and developer tools:

Ollama & Open-Weight Models: Used Ollama running locally on port 11434 to power local inference using open-weight models (localmind) and embedding generation (nomic-embed-text) with zero cloud latency or API billing.

Streamlit & PyMuPDF: Built the reactive desktop interface using Streamlit and handled local PDF document parsing, text chunking, and semantic vector indexing using PyMuPDF and NumPy.

MLH & DEV Ecosystem: Connected my Major League Hacking account profile via DevRelay CLI to track open-source contributions and share this build with the global community.

Hackathon Experience

Building NEXORA was an incredible journey combining systems engineering and practical AI deployment. Working from NIT Trichy, the atmosphere was charged with late-night coding sprints, debugging vector embeddings, and optimizing local model memory footprints. Collaborating with fellow hackers in the MLH community pushed me to focus heavily on practical developer utility, data privacy, and clean, high-performance UI design.

NEXORA — On-Device AI Workspace

ABHINAV — Sun, 25 Oct 2026 21:32:37 +0000

What I Built
I built NEXORA — On-Device AI Workspace, a fully offline, zero-cloud desktop application built to maximize personal productivity while letting you be gloriously lazy.

Instead of jumping between SaaS subscriptions, API key management, rate limits, and tab clutter, NEXORA aggregates model inference, document search, mathematical execution, and capital weather forecasting into a single unified dashboard.

It lets you submit long, dense PDFs or text documents and ask questions directly using local vector embeddings. You get back concise, page-cited answers without ever reading through endless pages or worrying about your private data leaving your computer.

Demo
Here is a glimpse of NEXORA running locally on a personal laptop:

accounts.google.com

Code
Below is the complete, single-file implementation for app.py:

app.py - Google Drive

drive.google.com

How I Built It
NEXORA relies on open-source software and open-weight models at every step:

Local LLM & Embedding Execution: Driven by Ollama, which runs an open-weight model (localmind) alongside open embedding weights (nomic-embed-text) locally on port 11434.

On-Device Vector RAG Engine: Document processing uses PyMuPDF (pymupdf) to extract and chunk text locally. Chunks are passed to nomic-embed-text to generate embeddings, and similarity scoring is calculated with NumPy (numpy) using cosine vector product distance.

Streamlit Desktop UI: Built using Streamlit, styled with a custom dark high-contrast theme (#0d1117) inspired by GitHub dark mode.

Why Does Open Innovation Matter?
Open innovation is the backbone of NEXORA. Using open-weight models locally offers crucial advantages:

Complete Data Privacy & Offline Capability: Sensitive financial, academic, or personal documents never leave your machine. You can analyze private contracts or read confidential papers on an airplane with Wi-Fi turned off.

Zero API Bills: Once downloaded, running open weights costs $0 per million tokens. You can summarize hundreds of documents without worrying about pay-per-token API charges.

Full Speed Transparency: Every response displays execution specs—total wall time, pure generation time, exact token generation counts, and generation speed (tokens/sec).

The Lazy Payoff
NEXORA embodies the "Slow and Steady" sloth mentality:

Zero Manual Reading: Drop a 50-page PDF, ask a question, and receive page-cited answers instantly.

No Internet Dependency: Works completely offline without needing cloud services or API connections.

Set Up Once, Run Forever: Spin up Ollama, run Streamlit, and your local workspace handles tasks automatically.

DEV Community: ABHINAV

NEXORA: Your Private, Local-First AI Assistant Powered by Open Source

NEXORA — On-Device AI Workspace

app.py - Google Drive