Killedar · किल्लेदार

Killedar · किल्लेदार

  1. Techology 🛠️
  2. 2 weeks ago
  3. 2 min read

Killedar (किल्लेदार) is the keeper of a Maratha fort. This one fits in your pocket and works where the network doesn’t.

The top of a Sahyadri fort usually has no signal, so an AI guide built on a hosted API stops working exactly where you need it. Killedar runs Gemma on the phone itself, so it keeps answering in airplane mode. It covers Sinhagad, Rajgad and Lohagad to start.

Built for the DEV Hacktoberfest 2026 Open-Source AI Challenge, Week 1: Touch Grass.

Demo

Watch the Killedar demo on YouTube

Features

  • Plan (works offline)

    • Pick a fort and a trail
    • Predicts your climb time from your past hikes (TabPFN, precomputed)
    • Works backwards from the last bus so you know when to leave Pune

  • Ask (works offline)

    • Questions answered on-device by Gemma via WebLLM
    • Grounded only in that fort’s notes
    • English or Marathi

  • Listen (works offline)

    • Fort history read aloud, with ElevenLabs narration bundled in the app or the phone’s own voice

Tech Stack

  • TypeScript PWA with zero runtime dependencies and a service worker for offline use
  • Gemma via WebLLM, running in the phone’s browser
  • TabPFN (Python) for climb-time predictions, precomputed into a 7 KB file
  • ElevenLabs for English and Marathi narration
  • Render for static hosting

How It Works

  • No backend at all. The app shell, fort notes, predictions and audio are cached by the service worker, and WebLLM caches the model weights.
  • TabPFN can’t run in a browser, but the app only ever asks about its own trails at a start hour (4 to 14) and group size (1 to 30). So all ~1,650 possible questions are answered ahead of time into a static JSON file.
  • Ports and adapters on both the web and Python side, with dependencies pointing inward.

Try It Out

On a phone: open the site in Chrome on Wi-Fi, tap Use offline, then Download for offline. Once the green check appears, switch to airplane mode and ask away.

Fort details come from cited public sources and are not yet verified on site, so please confirm timings locally before relying on them.

tech project ai offline pwa gemma