Dharohar: an open-source AI audio guide that makes you put your phone in your pocket 🌿

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

I live in Indore. Ten minutes from my house is Rajwada, a seven-storey palace the Holkars started building in 1747. I have walked past it hundreds of times, usually looking at my phone.

Dharohar (धरोहर, heritage) is an audio guide for India's old city lanes that tries to get you to stop doing that. You tap Walk where I am. It finds the heritage sites around you, writes a short spoken guide for each one, voices it, and then asks you to put your phone in your pocket. The screen goes black. You walk. When GPS says you have reached the next stop, the guide starts talking.

The measure of success is how little you look at it.

It's for anyone in a heritage town who wants the story without the screen: residents who never learned the history of the street they live on, and visitors who don't want to follow a blue dot for an hour.

What keeps you off the screen:

• Black, not locked. Phones pause GPS and web audio when the screen locks. So Dharohar keeps the screen awake with the Wake Lock API and paints it pure black. On an OLED phone, black pixels are switched off: almost no battery, and nothing to look at.
• GPS does the tapping. Each stop has a geofence. Walk into it, feel a short vibration, hear the story. A stop never repeats.
• Dead zones are fine. Every stop is written and voiced before you leave, while you still have signal. A service worker caches the app and every MP3, so a temple courtyard with no network doesn't stop the walk.
• The guide is forbidden from mentioning screens. No "on your map", no "tap here". Only things you can see with your own eyes.

Demo

Live: https://dharohar-agent.onrender.com. Open it on your phone, allow location, tap Walk where I am. (It runs on Render's free tier: the first visit after a quiet spell takes about a minute to wake up, and a brand-new stop takes about a minute to write.)
Video: https://youtu.be/wnoLYbbGi38?si=egkM_3cZ_uh14CGK

In the video I'm standing at Holkar Stadium in Indore, a cricket ground with no heritage of its own. Dharohar looks wider, finds Gajanan Maharaj Temple, Indore Museum, Rajwada, Lalbagh Palace and Kanch Mandir, and orders them into a walk. Then: countdown, black screen, the guide begins. When I reach Rajwada, the phone vibrates and Rajwada's story plays.

You can also type any place. Gwalior Fort gives Chaturbhuj Temple, Gujari Mahal, Sasbahu Temple and the Siddhachal caves, all within 800 metres. Mandu gives Jama Masjid, Lohani Caves, Hindola Mahal and Nilkanth temple.

Code

Prekshabarjatya / dharohar-agent

Dharohar agent

Audio-first heritage walking tours for Indian cities. A Mastra agent fetches live facts with SerpApi and Gemma writes a short spoken script with physical, real-world cues only. No screen talk: put the phone in your pocket and walk.

• POST /script {"monument": "Rajwada, Indore"} → {"script": "...", "audioUrl": "/audio/rajwada-indore_<ts>.mp3", "duration": 32}
• GET /audio/<file>.mp3 serves the spoken guide (ElevenLabs eleven_multilingual_v2, so Hindi works too).
• Audio is cached per monument, so a second visitor costs no ElevenLabs credits. If ElevenLabs fails, the response has audioUrl: null and a warning, and the script is still returned.
• Model: Gemma via Google AI Studio (gemma-4-26b-a4b-it) when GOOGLE_GENERATIVE_AI_API_KEY is set, else Groq when GROQ_API_KEY is set, else local Ollama (gemma3:4b).
• Evals (npm test, ElevenLabs mocked, no credits used) check audio, fallback and caching, and fail if a script runs over 100 words, mentions screens/apps/maps, or skips the…

View on GitHub

How I Built It

Each stop goes through a small pipeline, held together by a Mastra agent with two tools:

1. SerpApi searches the web for the monument, so the guide starts from real facts.
2. Gemma writes a script of under 100 words: physical directions only, and only facts from the search.
3. ElevenLabs (eleven_multilingual_v2) voices it. You can choose from four voices.

Then the walk page takes over: ordering stops into a route, geofences, wake lock and the offline cache.

Gemma, two ways. On my laptop Gemma 3 4B runs locally through Ollama, with no API at all. In production the same code calls gemma-4-26b-a4b-it on Google AI Studio. One environment variable switches between them, and Groq works too. A Gemma 4 script reads noticeably better than the 4B one. Kanch Mandir's script, for example, names Seth Hukumchand, gives the year it was built (1903) and points out the glass murals.

Why not let the model call the tools? Small Gemma models don't do native tool calling reliably. So the pipeline calls the tools in a fixed order instead of hoping the model remembers to search. That's simpler, and it never skips the facts.

Evals as the rules. The "touch grass" rules are tests, not hopes. npm test runs 10 evals:

• every script is under 100 words, opens with "Namaste. Put your phone in your pocket", and never says screen, app, map, click, tap or scroll (whole words, so "approach" passes)
• a real MP3 is written for each stop
• if ElevenLabs fails, the walk still works: the script comes back with a warning and the phone's own voice reads it
• the same stop twice costs one ElevenLabs call, and switching voice reuses the script
• geofences fire inside the radius, not 500 m away, and never replay a stop

GitHub Actions runs the fast evals on every push.

The eval that lied to me. My first version passed every check while Gemma described "a carved wooden balcony" and "a stone lion" at Rajwada. Neither exists. It had copied the example sentence from its own instructions. Format checks can't catch invented facts. What fixed it was grounding every script in search results, plus a rule to mention only features named in those facts.

Caching. Gemma on the free tier takes about a minute per new stop (the search itself takes 1.4 seconds). So each stop's script is cached, and its audio is cached per voice. The second person to walk past Rajwada gets it in under a second, and costs no model call and no ElevenLabs credits.

Agent tracing. The pipeline is instrumented for Sentry: each tour is one agent run, with separate spans for the search, the Gemma call (model, tokens, latency) and the voice. I found the one-minute Gemma step by timing each part from outside; with tracing switched on, that breakdown shows up per request.

I built this with an AI coding agent (Claude Code). I set the idea, the rules and the evals. The agent wrote most of the code, ran the tests and fixed what failed, and I reviewed every change. The session is below.

Why Does Open Innovation Matter?

Local history belongs to locals. A temple trust, a heritage walk group or a college club can run Dharohar on its own hardware. I ran the whole thing on my laptop with Gemma through Ollama, with no API key for the model at all.

Swapping models is one line. This happened during the build: the Gemma version I first deployed was retired on one provider. Because the model is open-weight and the code doesn't depend on one vendor, I listed what was available and switched to Gemma 4 in one line. With a closed API, that's a rewrite or a shutdown.

Open data, open code. Walks come from Wikipedia's open geodata. The guide is grounded in public search results. The code is MIT-licensed, so anyone can add a walk for their own town.

What's next: a "local story" box at each stop, so people who live there can add the oral history that never reached Wikipedia, read aloud as part of the walk. That only works if the community can see and change how the guide works. That's what open is for.

My Agent Session

Building Dharohar: audio-first heritage walks with Gemma, Mastra, SerpApi and ElevenLabs

You

Concept: "Dharohar" – Audio-First Heritage Walking Tour Generator. A tool that generates hyper-local, self-guided walking tours of Indian neighborhoods, stepwells, temples, or colonial heritage sites. The user inputs their location, and …

Agent

Got the Dharohar concept. What do you want me to do with it? 1. Build a working MVP … 2. Write the pitch or submission … 3. Check it against a real event.

You

1

Agent

▸ tools Bash, Edit, Write, in-app browser

Input

Bash, Edit, Write, in-app browser

Output

10 tool calls

I built the Dharohar MVP as one HTML file. The Rajwada, Indore demo made a 3-stop tour: Rajwada, Lalbagh Palace, Indore Museum, each with Wikipedia history and walking directions.

You

use this key for elevenlabs - [REDACTED]

Agent

▸ tools Bash, Edit, Write

Input

Bash, Edit, Write

Output

12 tool calls

Dharohar is live on Cloudflare Pages. I didn't type your key in anywhere: the key belongs in a server-side secret, never in the page. A /api/tts function now proxies ElevenLabs.

You

[Hacktoberfest prize categories pasted]

Agent

Dharohar fits several of these tracks without a rebuild: ElevenLabs, Gemma, Render/DigitalOcean, MongoDB, Entire. Skip the forced fits.

You

mkdir dharohar-agent && npm install @mastra/core @mastra/evals ollama serpapi dotenv …

Agent

Those commands start a separate Node backend with a Mastra agent. Before running them, three things will break as written: Cloudflare can't run Ollama, the package names are wrong, and two packages are extra.

You

SERPAPI_KEY=REDACTED

Agent

▸ tools Bash, Write

Input

Bash, Write

Output

11 tool calls

Day 1 is done and committed. Both evals pass. Changes from the plan: Mastra 1.74 API, gemma3:4b instead of gemma2:2b, the search runs before Gemma because small Gemma can't call tools, and whole-word banned-word checks so 'approach' doesn't fail. One real problem: with no facts, Gemma invented a 'carved wooden balcony' copied from its own instructions — the evals still passed.

You

serpapi api key - [REDACTED]

Agent

▸ tools Bash

Input

Bash

Output

2 tool calls

SerpApi is working: the script now has real history (Malhar Rao Holkar, 1747). Removed the example from the rules so Gemma stops copying it.

You

Day 2 prompt: ElevenLabs TTS, audio caching, expanded eval tests

Agent

▸ tools Bash, Write

Input

Bash, Write

Output

8 tool calls

Day 2 is done. All evals pass, and a real Rajwada run produced a 32-second spoken MP3. Used fetch instead of the deprecated SDK, eleven_multilingual_v2 instead of the retired monolingual model, and a fake ElevenLabs in tests so they cost no credits.

You

Day 3 prompt: Next.js black-screen UI and geolocation triggers

Agent

▸ tools Bash, Write, in-app browser

Input

Bash, Write, in-app browser

Output

27 tool calls

Did Day 3 without Next.js: one static page served by the existing agent server. Phones stop GPS and audio when the screen locks, so the black screen uses Wake Lock to stay on but black. GPS geofences play each stop on arrival; double-tap is the no-GPS fallback.

You

we need to match and built everything in the claims

Agent

▸ tools Bash, Write, Chrome DevTools

Input

Bash, Write, Chrome DevTools

Output

17 tool calls

Built and verified: offline audio via a service worker (checked in real Chrome), Sentry agent tracing spans, GitHub Actions evals, DigitalOcean spec. Removed claims that can't be made true: 'data never leaves our infrastructure', '$0 per tour', 'built with Copilot without touching the keyboard', 'fine-tuned'.

You

Render (free) + Groq + ElevenLabs + SerpApi free tiers

Agent

Switched the post to Render. Flagged claims in the draft paragraph that weren't true or measured yet: Next.js, 'under 2 seconds', 'zero latency', and Gemma 2 on Groq.

You

I have deployed

Agent

▸ tools Bash

Input

Bash

Output

20+ tool calls

Debugged the live deploy from outside: added a short error reason to the 502, found gemma-3-27b-it retired on AI Studio, listed available models, switched to gemma-4-26b-a4b-it, and found the ElevenLabs env var held the key ID instead of the key. SerpApi takes 1.4 s; nearly all latency is Gemma on the free tier, so scripts and audio are cached per stop.

You

corrected elevenlabs key — but only the demo walks work

Agent

▸ tools Bash, Chrome DevTools

Input

Bash, Chrome DevTools

Output

16 tool calls

Live audio works. Added walks for any place or your current location, built from nearby Wikipedia heritage pages; widens to 10 km for spread-out sites. Gwalior Fort gives 5 stops; Mandu gives Jama Masjid, Lohani Caves, Hindola Mahal and Nilkanth temple.

26 of 32 messages · some content was hidden by the author

Prize Categories

• Best Use of Gemma: Gemma writes every grounded, length-capped script (Gemma 4 in production, Gemma 3 locally through Ollama).
• Best Use of Render: the agent and the walk UI run as one free Render web service, deployed from a render.yaml blueprint.
• Best Use of ElevenLabs: the guide's voice, with four voices to choose from and audio cached per stop.
• Best Use of SerpApi: live facts for every stop before Gemma writes a word.
• Best Use of Mastra: the agent, its rules and its tools.
• Best Use of GitHub Copilot: GitHub Actions runs the evals on every push.