Lumina
Team led by a YC founder and CTO alongside a Waterloo CS founding engineer, specializing in on-device LLMs, agentic RAG, computer vision, and C++/Python.
YouTube Video
Project Description
Lumina is the first private, instant, agentic tutor that runs 95 % on-device for yound students in rural and low-connectivity regions of the Global South. In under 8 hours at the DeepMind × Cactus Hackathon, we built a React Native app where a child simply snaps a photo of handwritten or records a voice note of his homework (fractions, decimals, early algebra) and receives personalized step-by-step tutoring in <800 ms — completely offline, with zero data leaving the phone.
Flawless on-device execution is powered by FunctionGemma-270M-it running inside the Cactus Compute runtime. FunctionGemma handles the entire agentic workflow locally: vision-based problem extraction, multi-step reasoning (plan → solve → reflect), tool calling (local math solver + canvas drawer for visual explanations), practice question generation, and persistent personalized memory via Cactus local vector RAG. This delivers lightning-fast, private experiences even on mid-range $80–120 Android phones with no internet.
Cactus Compute is the critical runtime layer that makes sophisticated local agents production-ready. It provides reliable agent orchestration, long-running workflow management, and the intelligent routing logic that decides in real time: 92–95 % of interactions stay 100 % on-device for maximum speed and privacy; only complex reasoning tasks (e.g. mathematical proofs) trigger a seamless, streaming fallback to Gemini 2.0 Flash. The handoff is invisible to the user except for a tiny transparent toast, keeping conversation continuity perfect.
This architecture fundamentally relies on FunctionGemma for fast local function calling and agent loops, and Cactus Compute for the hybrid edge-cloud infrastructure. Together they redefine what’s possible at the edge-cloud frontier, proving that true agentic AI education can finally reach the 2 billion children the cloud cannot reliably serve.
Technologies used: React Native (Expo scaffolding), cactus-react-native SDK, FunctionGemma-270M-it (INT4 quantized via Cactus), Gemini 2.0 Flash API (fallback), Cactus local vector store + SQLite, custom Cactus tools (math solver, drawing canvas).