DEV Community

Faheem Musthafa P
Faheem Musthafa P

Posted on

Spot-Undoo v2: Open-Weight AI Layover Studio & GPU Map Engine for Kerala

#devchallenge #hf26challenge #ai #opensource

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 3.

🌟 Spot-Undoo v2: Open-Weight AI Travel Studio & Layover Optimizer

Spot-Undoo is an open-source, privacy-preserving micro-travel concierge engine tailored for travelers, transit layover passengers, and explorers across Kerala's 14 districts.

During Week 3, we advanced Spot-Undoo from a mobile-first app into a full-scale Open-Source AI Desktop Studio & GPU-Accelerated Geo-Engine, designed to deliver instant, constraint-guaranteed transit itineraries with zero cloud inference cost.

🚀 What's New in Week 3 Development

  1. 🖥️ Widescreen Desktop Studio Layout (1680px):

    • Complete architectural redesign away from cramped mobile layouts into a split-screen studio workstation.
    • Docked Inspection Panel: Left-docked details card with high-resolution photography, transit breakdowns, and authentic community reviews.
    • Expanded Canvas: Right-hand interactive Leaflet map canvas utilizing full desktop screen real estate.
    • Dual-Column Layover Assistant: Dedicated left-side chat cockpit with right-hand live telemetry inspector.

  2. ⚡ GPU-Accelerated Leaflet Map (Zero Pan Lag):

    • Eliminated panning and zoom stutter across large Kerala spot clusters using memoized zoom tiers and batched L.layerGroup.
    • Hardware-accelerated CSS transforms (will-change: transform; transform: translate3d(0,0,0)) achieving smooth 60fps rendering.

  3. 🧠 Open-Weight Local AI Inference (Ollama & Qwen):

    • Powered by local open-weight models (Qwen 2.5:3B / Gemma 4) with structured JSON outputs.
    • Guaranteed constraint matching (max available time, strict budget limits, transit travel time).
    • Instant mathematical fallback routing if running completely offline.

  4. 📸 On-Device Multimodal Verification:

    • HTML5 Canvas client-side downsampling to 512px JPEG before sending to vision models, keeping memory footprint low.
    • Instant visual spot analysis with vibe tagging and authentic local tips.

🛠️ Open-Source Tech Stack

  • Frontend: React 18, Vite, Tailwind CSS 3, Lucide Icons, Leaflet OpenStreetMap.
  • AI Models: Local open-weight models via Ollama (qwen2.5:3b, gemma4), Groq Cloud API, Google Gemini Flash.
  • Maps & Data: OpenStreetMap, Leaflet GPU acceleration, LocalStorage zero-database architecture.
  • Mobile & Android: Capacitor native Android bridge (SpotUndoo-v1.0.apk).

🔗 Open-Source Repository & Demo

Built with ❤️ for Hacktoberfest 2026 & MLH Hackathons.

Top comments (0)