Skip to Main Content

galaxy

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

galaxy — Demo video

SIGHTLINE - Point your phone at anything, and it shows you exactly where to touch. The problem: every day, millions of people stand in front of a blinking router, a washing machine flashing E03, or a bank KYC form, and freeze. The manual is in technical English. A chatbot asks them to describe what they see, but they don't know what the thing is called. You can't search for a button you can't name. What SIGHTLINE does: you take a photo and ask, in English or Malayalam. SIGHTLINE answers with a numbered sequence of steps, and for each step it lights up the exact physical part - the WAN port, the reset pinhole, the drain filter door, the signature box - with a pixel-precise mask right on your photo, then reads the step aloud. No jargon, no guessing, no technician visit for a two-minute fix. How it's built - two Meta models in one loop: 1. Meta Llama 3.2 11B Vision (Cloudflare Workers AI) is the brain. It reads the image and the question, reasons about the device's state, writes the steps in the user's language, and grounds each step to a bounding box. 2. Meta SAM 2.1 (local, via Ultralytics) is the eyes. It turns each rough box into a pixel-perfect mask so the user sees the exact button, not a vague rectangle. 3. A Streamlit UI with step-by-step navigation, plus gTTS voice in Malayalam and English. Why it's different: most vision apps stop at image-to-text. SIGHTLINE goes image-to-action. Llama's reasoning feeds SAM's segmentation, turning language into a precise visual pointer that neither model can produce alone. Challenges: with hours left and no API credit card, we rebuilt the whole pipeline on a fully open, free Meta stack - runnable by anyone, anywhere, at zero cost, which is exactly what our users need. Making Llama's coordinates line up with SAM required defensive JSON parsing, coordinate clamping, and graceful text-only fallbacks so the app never crashes. What's next: live camera mode with SAM 2.1 video tracking, more Indian languages (Hindi, Tamil, Kannada, Telugu), offline on-device inference, and hands-free guidance on Ray-Ban Meta glasses, where SIGHTLINE's masks become an AR pointer in your field of view. Solo Indie Track entry by Abhishek A.S., India.

Team