North Star (Polaris)

Offline-first, local-LLM study and fitness agents that run entirely on your own device — no API keys, no account

All projects  ·  About Matt

A submission to the North Star Challenge (Access & Equity track). Built with AI-assisted development — designed and directed by me, implemented with AI pair programming.


Description


What it does: North Star is a set of downloadable, privacy-preserving AI components that run entirely on a user's own device — no API keys, no account, no internet required. Everything is orchestrated with LangGraph state machines and powered by a local LLM through Ollama, so a student with a laptop and no data plan still gets a full study and fitness toolkit.

Why it matters: The whole project is built around access and equity. It's designed for a group of students in one room with no internet — often sharing one device — so it ships features like Study Packs (bundle decks and quizzes into a portable file, shared by USB or messaging, no server) and a Group Quiz mode for studying together offline.

How it was built: A Python monorepo with a shared polaris_core (LLM, embeddings, config, memory) and four installable packages. LangGraph routes each request to the right specialist; Chroma stores notes locally for cited retrieval; the fitness pipeline is defined by editable markdown agent files. A serverless hosted version also exists (React SPA + FastAPI on Cloud Run with a Groq fallback) for people who can't run a model locally.

The four components


North Star (Polaris) — Four Offline Components ═════════════════════════════════════════════════ ┌──────────────┬────────────────┬────────────────┬────────────────┐ │ Study LLM │ Study RAG │ Fitness Agents │ College Planner│ │ (6 areas) │ (vector notes) │ (MD-defined) │ (app tracker) │ │ │ │ │ │ │ Flashcards │ ingest → embed │ parse .fit / │ deadlines / │ │ Quizzing │ → store → │ .tcx / .gpx │ status / notes │ │ CV Builder │ retrieve → │ → metrics → │ 4-year credit │ │ Advisor │ grade → │ growth plan → │ map │ │ Citations │ cited answer │ review │ export → .ics │ │ Essay Helper │ (Chroma, disk) │ (agent mds/) │ │ └──────────────┴────────────────┴────────────────┴────────────────┘ shared core: polaris_core (LLM · embeddings · config · memory) LangGraph + Ollama (all local)

Dev Notes

Problem Solved

Most study AI assumes reliable internet, an account, and money for API calls. North Star removes all three — it downloads once and runs on the device, so cost and connectivity stop being barriers to good study tools.

Technical Highlights

Four LangGraph-orchestrated components share one local core. The Study LLM routes across the six Polaris areas; Study RAG gives cited answers from a local Chroma DB; fitness agents are pure markdown you can edit without touching code; and Study Packs make everything shareable offline.

What's Next

Native Android and iOS wrappers (scaffolds are in the repo), a smoother one-command installer, and expanding the College Planner's 4-year credit-mapping.


View the submission on GitHub


Back to top