DEV Community: Harshit Anand

Trailtale: A Pocket-Sized Audio Story That Gets You Off Your Screen and Back on the Trail

Harshit Anand — Sun, 11 Oct 2026 20:17:41 +0000

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

What I Built

Trailtale is a local, offline walking app that turns a walk into an immersive audio story. While you move, the app plays a short story segment; when you stop, the audio pauses automatically. It also surfaces weekly nature quests and lets you confirm interesting finds with a photo, all without pulling you back to a screen.

It’s built for people who want to spend more time outside without getting stuck in their phone. The idea is simple: make the walk the main experience, not a distraction from it.

Code

Repository:

HarshitAnand1 / trailtale-touch-grass

Trailtale

An audio story that plays only while you walk, plus weekly nature quests. Every model runs locally — no cloud, no account, works offline.

Quick Start

1. Install Python dependencies

cd server
pip install -r requirements.txt

2. Start the server

# From the project root
uvicorn server.main:app --host 0.0.0.0 --port 8000 --reload

3. Open the app

Visit http://localhost:8000 on your phone (or laptop browser).

4. (Optional) Set up llama.cpp for story generation

# Download a Gemma model and start the llama.cpp server
./llama-server -m gemma-2b.gguf --port 8080

Without llama.cpp, the story engine returns fallback text so you can still test the full walk flow.

5. (Optional) Install Piper for TTS

# See https://github.com/rhasspy/piper for installation
piper --model en_US-lessac-medium --output_file test.wav <<< "Hello from the trail"

Without Piper, the TTS engine generates silent placeholder audio.

Architecture

Phone (PWA)              Host (FastAPI)
┌──────────────┐        ┌──────────────────┐
│ Walk Gate    │──────▶ │ Orchestrator

…

View on GitHub

Project source: this includes the PWA client, FastAPI backend, quest data, and local AI pipeline.

How I Built It

I built Trailtale as a combination of a mobile-first PWA and a local AI backend:

Frontend: a lightweight browser app in the client/ folder using HTML, CSS, and JavaScript
Sensors: phone motion and GPS data are used to detect whether the user is walking or stopped
Backend: FastAPI in the server/ folder orchestrates the walk flow, story queue, quests, and logs
Story generation: a local model runs through llama.cpp using Gemma
Voice: Piper generates spoken audio from the generated story text
Quest system: weekly nature quests are stored as local JSON packs and tracked in SQLite
Optional vision check: a local multimodal model can verify that a photo matches a quest item
Offline-first design: the app is intended to work with no cloud service, no account, and no internet dependency

This stack makes the app feel like a real walk companion rather than a web app that happens to use AI.

Why Does Open Innovation Matter?

Open innovation mattered here because the whole product depends on privacy, local control, and flexibility.

A closed API would have forced:

cloud-based story generation for every segment
network calls for TTS and photo verification
higher cost per walk
user data leaving the device

With open-source tools and open-weight models, Trailtale can run locally on a phone, laptop, or a backpack setup. That means:

no cloud dependency
no account required
no pay-per-token barrier
the prompt, model, and voice can all be tuned or swapped easily

For an app designed to get people to step away from screens, local open tech is not just a technical choice—it’s the product requirement.

Prize Categories

Gemma

You Said You'd: A 100% Local AI Commitment Tracker for Forgetful Friends

Harshit Anand — Sun, 04 Oct 2026 23:44:19 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

*"You Said You'd" * is a local, privacy-first AI commitment tracker.

I built this for a friend who is incredibly busy and constantly making casual promises in WhatsApp and Telegram groups ("I'll send you that link tomorrow", "Let's grab lunch next week", "I owe you $20 for the tickets"). Unsurprisingly, these casual commitments get lost in the noise of group chats, leading to forgotten plans and missed IOUs.

To solve this, I built a brutalist, retro-themed web application where they can drop in a raw chat export (.txt or .json). The application parses the chat and uses local AI to extract every single promise, request, and IOU made by them or to them, categorizing them into a neat, searchable ledger.

When I handed the app over to him and had him upload his group chat from last month, his exact reaction was: "Wait, I completely forgot I told Shivam I'd look at his resume. This is terrifyingly useful."

Demo

hacktoberfest_challenge_1.mp4 - Google Drive

drive.google.com

Code

HarshitAnand1 / local-commitment-tracker

local, privacy-first AI commitment tracker

You Said You'd

An AI that remembers what you promised, what others promised you, and shows the exact message where it was said.

Hacktoberfest Weekend Challenge 2026 · Build for a Friend + Open-Source AI

The Problem

People constantly make small commitments in chats:

"I'll send it tonight."
"I'll bring the HDMI adapter."
"I'll talk to the professor tomorrow."
"Don't worry, I'll handle the registration."
"I'll send you the deck after dinner."

These commitments disappear into thousands of messages. You Said You'd imports a user's own chat export and turns those vague conversational promises into a searchable, evidence-backed commitment ledger.

This is not another generic chatbot. Its job is specifically:

Remember what I said I'd do. Remember what other people said they'd do. Help me resume unfinished commitments.

Target User

A real person who frequently forgets small promises, has many active projects, works in groups/clubs, loses track of who owes…

View on GitHub

How I Built It

The entire stack is designed to run entirely locally, turning your own machine into the AI server:

Frontend: Next.js (App Router) styled with vanilla CSS following a strict Hacktoberfest 26 retro-brutalist theme (no Tailwind).
Agent Framework: I used Mastra (@mastra/core) to orchestrate the AI logic. Mastra handles the multi-step workflow of parsing the chat and enforcing structured JSON outputs.
Local Inference: Ollama running Gemma 3 natively on the machine. This acts as the brain for the Mastra Agents.
Memory & Retrieval (RAG): Instead of dumping massive chat histories into the context window, I used PostgreSQL with pgvector. Every commitment is embedded using the nomic-embed-text model. When the user asks a question on the "Ask" page, we do a cosine similarity search to find the top 5 most relevant commitments and pass only those to the Mastra QA Agent.

Why Does Open Innovation Matter?

This project is built around one core philosophy: My friend's conversations are not AI training data. They're their private life.

A person's chat history contains personal relationships, private plans, arguments, addresses, and sensitive financial context. It is incredibly dangerous to send 5 years of WhatsApp history to a closed-source cloud LLM API just to extract some reminders.

Open innovation—specifically the availability of powerful open-weights models like Gemma 3, local inference runners like Ollama, and open-source frameworks like Mastra—made this project possible. I was able to build an advanced RAG pipeline that processes sensitive data entirely on the user's local hardware. You get the power of an AI assistant while keeping your data off a server you don't control.

Prize Categories

Best Use of Gemma (Ran Gemma 3 locally as the core reasoning engine for extracting commitments).
Best Use of Mastra (Orchestrated the extraction workflow and the QA agent using Mastra over open local models).