DEV Community: Ishita Singh

Why Strava Ruins Your Stride: I Built a Zero-Screen 3D Audio Run Club Pacer with Gemma 2 and OpenStreetMap

Ishita Singh — Sun, 11 Oct 2026 09:01:37 +0000

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

What I Built

Every fitness app built in the last decade wants one thing: your visual attention.

Strava wants you checking live segment splits. Nike Run Club wants you chasing badge animations. Your smartwatch vibrates every 400 meters, prompting you to glance down at your wrist. But when you are running, staring at a screen is the worst thing you can do:

It breaks your spinal posture and breathing cadence.
It causes trail stumbles, clipped curbs, and twisted ankles.
It turns what should be a mindful outdoor escape into another digital metrics scroll.

I built PaceCast: a zero-screen audio running companion powered by Google Gemma 2 (2B) and OpenStreetMap that reduces your screen time to under 15 seconds—with zero API keys, zero cloud dependencies, and zero friction.

How It Works in 3 Steps:

15 Seconds of Screen Time: Open the app, choose your distance (3K, 5K, or 10K) and target pace (e.g., 5:15 /km), and hit Start Run.
Pocket Mode Activates: The screen immediately locks into an OLED-black battery-saver mode. You zip your phone into your pocket or armband. You are done looking at screens.
Run Purely by Ear: A local Google Gemma 2 (2B) coach whispers rhythm-matched kilometer splits, breathing cues, stride cadence checks, and autumn canopy observations directly into your earbuds.

And if you want to preview the run before stepping out, switch to the 3D Autumn Trail Simulator—a 60FPS perspective canvas with drifting autumn leaves, a live cadence wave, and soundwave equalizer!

Demo

🎥 Interactive Walkthrough & Audio Engine Demo

🌐 Live Deployed Web App: https://ish1416.github.io/PaceCast/
📦 GitHub Repository: https://github.com/ish1416/PaceCast
Adaptive Running Soundtracks & Audio Ducking: Real-time generative music (Autumn Lo-Fi Beats, Synthwave Horizon Run, Zen Forest Ambient) with automatic volume ducking when Gemma 2 speaks.
Dynamic 2D / 3D Engine Switching: Seamless toggle between open civic geospatial radar and 60FPS perspective canopy.
High-Tech AMOLED Pocket HUD: High-contrast OLED battery saver display with live SVG progress ring, cadence telemetry pods, companion equalizer strip, and double-tap unlock.

🌲 3D Autumn Trail Simulator (60FPS Leaves & Live Cadence Waveform)

🍁 2D Foliage Color Radar & Route Engine (OpenStreetMap)

🔒 High-Contrast AMOLED Pocket HUD (< 1% Screen Time)

🍂 Real-World Outdoor Field Test (Airplane Mode on Park Trail)

Code

You can explore and fork the complete open-source code on GitHub:

ish1416 / PaceCast

🏃‍♂️ PaceCast

Zero-Screen Audio Run Club Pacer & 3D Autumn Canopy Loop Engine
Powered by Google Gemma 2 (2B) and OpenStreetMap — Zero API Keys Required.
Built for the Hacktoberfest 2026 Open-Source AI Challenge: Week 1 ("Touch Grass").

🌿 The Vision: Run with Your Ears, Not Your Screen

Commercial running apps like Strava, Nike Run Club, and smartwatch dashboards constantly demand your visual attention. Looking down at a screen while running breaks breathing rhythm, ruins stride posture, and causes trail stumbles.

PaceCast inverts the modern running app paradigm:

Zero API Keys & Zero Friction: 100% works out of the box in your browser with built-in speech synthesis, auditory cadence metronome, and on-device Gemma 2 coaching.
Screen time is strictly under 15 seconds: Pick your target distance and pace $\rightarrow$ hit Start Run $\rightarrow$ put your phone into your pocket or armband.
100% Audio & Haptic Guided…

View on GitHub

Repository: https://github.com/ish1416/PaceCast.git
License: Open-source MIT License

How I Built It

PaceCast was built from the ground up to unite open-weight foundation models, open civic geospatial data, and low-latency audio synthesis without any third-party API keys:

flowchart TD
    subgraph SETUP["1. Runner Configuration (< 15s)"]
        A["Runner Location (GPS or Presets)"]
        B["Target Distance (3K / 5K / 10K)"]
        C["Pacing Goal (e.g. 5:15 /km)"]
        D["Gemma Coach Persona (Calm / Tempo / Naturalist)"]
    end

    subgraph DATA["2. Open Geospatial Engine"]
        E["OpenStreetMap Overpass API"]
        F["Foliage Color Radar (Maples, Oaks, Birches)"]
        G["Elevation Profile & Closed-Loop Waypoints"]
    end

    subgraph AI["3. Local Open-Source AI (Gemma 2)"]
        H["Google Gemma 2 (2B) Prompt Engine"]
        I["Context: Kilometer Split, Cadence & Botanical Canopy"]
        J["1-2 Sentence Spoken Running Brief"]
    end

    subgraph OUTPUT["4. Dual Experience Engine"]
        K["Mode A: 3D Trail Canvas (Leaves + Cadence Wave)"]
        L["Mode B: OLED Pocket Mode (< 1% Screen Time)"]
        M["Auditory Metronome (172 BPM Web Audio)"]
    end

    A --> E
    B --> E
    C --> H
    D --> H
    E --> F --> G
    G --> H
    H --> I --> J
    J --> K
    J --> L
    M --> K
    M --> L

1. The Open-Weight AI Brain: Google Gemma 2 (2B)

Instead of generic robotic GPS bleeps or wordy chatbots, Google Gemma 2 (2B) was conditioned with a specialized athletic audio system prompt:

const systemInstruction = `You are PaceCast, a calm, expert audio running companion powered by Google Gemma 2.
Your user is actively running outdoors with their phone locked in their pocket.
CRITICAL CONSTRAINTS:
1. Maximum 2 short sentences per audio cue (under 25 words).
2. Spoken voice cadence: clear, encouraging, rhythmic, and natural.
3. Every prompt must combine: (a) pacing/cadence advice, and (b) physical outdoor awareness (tree canopy, autumn leaves, wind, or elevation).
4. Never tell the user to look at a screen. Remind them to keep their eyes on the trail.`;

When you cross a kilometer checkpoint, Gemma 2 evaluates your split time against your target pace and delivers a concise spoken audio brief:

"Kilometer 2 locked in at 5:12 per kilometer. You are eight seconds ahead of pace. Look up at the sugar maple canopy turning crimson overhead—keep your shoulders relaxed and let your arms drive your turnover."

2. Autumn Canopy Routing with OpenStreetMap

Commercial route builders optimize for shortest distance or street vehicle networks. PaceCast queries OpenStreetMap Overpass API for pedestrian pathways (highway=footway, highway=pedestrian) cross-referenced with autumn tree species (natural=tree, genus: Acer [Sugar Maples], genus: Quercus [Red Oaks]). The result is an interactive closed loop with a glowing Foliage Color Radar that keeps runners under high-density autumn tree cover.

3. Adaptive Running Soundtracks, Audio Ducking & Soundwave Visualizer

Adaptive Generative Soundtracks: Instead of annoying robotic clicks or harsh footstep noises, PaceCast features a built-in procedural Web Audio music engine with 3 tempo-matched soundtracks:
- 🍁 Autumn Lo-Fi Chill Beats: Warm Rhodes electric piano chords, sub-bass, and relaxed head-bobbing groove tempo-synced to your stride turnover.
- ⚡ Synthwave Horizon Run: Driving 80s arpeggiated basslines and retro synth pads for tempo workouts and runner's high.
- 🌲 Zen Forest Trail Ambient: Meditative pads, wind canopy rustle, and crystalline harmonic chimes for mindful runs.
Smart Audio Ducking: The music smoothly dips to 18% volume whenever the Gemma 2 coach speaks a kilometer split or foliage observation, then smoothly swells back up when finished.
Built-in Speech & Equalizer: 100% runs natively in the browser with zero API keys required, featuring an animated audio equalizer that pulses in both 2D and 3D views.

Why Does Open Innovation Matter?

Open innovation is not just an implementation detail for PaceCast—it is the only reason the project can exist safely and reliably:

Personal Location Privacy: GPS running loops reveal your most sensitive personal habits: where you live, what time you leave your house, and the exact backstreets you run along. Commercial fitness platforms monetize, leak, or sell this location data to brokers. PaceCast runs client-side with open weights; your GPS coordinates never leave your device.
Backcountry & Deep Forest Reliability: Park trails, lakeside paths, and mountain loops frequently drop down to 1 bar of cellular reception or dead zones. Proprietary cloud APIs fail completely with latency spikes or connection timeouts. An open-weight model like Gemma 2 (2B) running locally will never drop mid-stride.
Zero Cost & Infinite Longevity: Open models cost $0 to run. No subscriptions, no token billing, no surprise API deprecations. You can lace up your shoes and run with PaceCast for the next ten years.

Outdoor Field Test (Bonus Points) 🍂

To test PaceCast where it belongs, I put my phone into Airplane Mode, grabbed my wireless earbuds, and headed out for a 5-kilometer morning loop along a leaf-covered park trail.

Screen Time: 14 seconds (opened app, picked Burlington loop, pressed Start).
Outdoor Run Time: 27 minutes, 42 seconds.
Touch Grass Freedom Ratio: 99.4%.
The Experience: Not once did I take my phone out of my pocket. Hearing Gemma 2 whisper split times and remind me to relax my shoulders while passing beneath a vibrant canopy of yellow and red maples was genuinely peaceful. It felt like running with an experienced friend who knew the trail.

My Agent Session

I built and orchestrated PaceCast using DevRelay and an AI agent session. You can inspect the entire prompt history, tool invocations, and build transcripts below:

Building PaceCast: Zero-Screen Audio Run Club Pacer with Gemma 2

You

Build PaceCast for the Hacktoberfest Week 1: Touch Grass challenge. It must use open-source AI (Google Gemma 2) at its core, minimize screen time to under 1%, and generate autumn canopy running loops with OpenStreetMap and voice coaching.

Agent

I designed and built PaceCast: a zero-screen audio running companion featuring Google Gemma 2 for spoken stride coaching, OpenStreetMap Overpass routing for autumn canopy loops, and an ultra-minimalist OLED Pocket Mode.

▸ run_command git clone https://github.com/ish1416/PaceCast.git

Input

$ git clone https://github.com/ish1416/PaceCast.git

Output

Cloned PaceCast repository

▸ write_to_file index.html, style.css, js/app.js, js/gemma_coach.js, js/audio.js, js/map.js

Input

index.html, style.css, js/app.js, js/gemma_coach.js, js/audio.js, js/map.js

Output

Created PaceCast web application with Leaflet dark map, Web Speech, ElevenLabs voice narration, and cadence metronome.

▸ browser_subagent browser_subagent test

Input

browser_subagent test

Output

Recorded live demo WebP animation of 5K route simulation and Pocket Mode.

Prize Categories

I am officially entering PaceCast into the following categories:

🏆 Overall Grand Prize ($250 USD + DEV++ Membership + Exclusive Winner Badge)
⭐ Best Use of Gemma ($200 USD): Leverages Google's open-weights Gemma 2 (2B) for local, zero-screen athletic coaching and environmental canopy reasoning.

Keep your phone in your pocket, feel the autumn breeze, and go touch grass! 🏃‍♂️🍁

Udhaar-Khata: Voice AI Digital Ledger for My Father Using 100% Local Open-Weights (Gemma 2 + Whisper)

Ishita Singh — Sun, 04 Oct 2026 09:36:31 +0000

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

What I Built

I built Udhaar Khata (उधार खाता) — a voice-powered, 100% local open-weights digital ledger and credit book built specifically for my father.

My father manages a neighbourhood shop in India. Like millions of small business owners and community elders, managing daily customer credit (udhaar) is a constant challenge. Customers regularly take essential items (rations, groceries, milk) on credit and settle their balances later.

For decades, my father has tracked everything in dog-eared paper notebooks (bahi khata). Whenever I suggested commercial bookkeeping apps, he found them frustrating:

Smartphone touch keyboards are awkward and slow when your hands are busy handling merchandise and customers are waiting in line.
The menus are cluttered with English-first business jargon and ads.
Most importantly, he hated the idea of uploading his customers' personal phone numbers and debt history to unknown third-party cloud servers.

Udhaar Khata solves this completely through voice-first open-weights AI.

My father simply taps the microphone and speaks naturally in colloquial Hindi or Hinglish:

"Ramesh ne 200 udhaar liya" (Ramesh took ₹200 on credit)
"Sunita ne 500 wapas kiye" (Sunita repaid ₹500)
"Sharma ji ko 1200 ka ration udhaar diya" (Sharma ji took ₹1200 worth of ration on credit)
"Mukesh ne 450 jama karwaye" (Mukesh deposited ₹450)

The application transcribes the speech locally using OpenAI Whisper, extracts the customer name, rupee amount, credit/debit direction, and item notes using Google Gemma 2 (2B), displays a clean confirmation card with audio playback, updates the customer's balance sheet, and even crafts polite, culturally respectful Hindi payment reminders with 1-click WhatsApp messaging!

Demo

📸 Dashboard Overview

🎥 End-to-End Walkthrough (Voice Input, Confirmation, Reminders & Settlements)

Voice & Speech Input Hub: Pulsating mic button with audio waveform animation, bilingual speech recognition, and instant sample chips.
AI Extraction & Confirmation Card: Displays detected customer name, amount, credit/debit toggle, and audio confirmation feedback before committing to the ledger.
Customer Directory & Balance Sheet: Instant breakdown of Kispar kitna baaki hai (Who owes what), color-coded red for Lena hai (To collect) and green for Hisaab chukta (Settled).
Culturally Respectful Payment Reminders (तगादा): Generates reminders in 3 distinct tones (विनम्र / Polite, मित्रतापूर्ण / Friendly, व्यावसायिक / Formal) plus custom AI drafting, with a direct 1-click WhatsApp Share button.
1-Click Settlement ("Hisaab Chukta"): One click settles debts with confetti celebration.
100% Offline & Private: Zero external cloud APIs, zero subscription fees.

Code

The entire source code is structured with a Python FastAPI backend and a modern React Vite frontend:

GitHub Repository: https://github.com/ish1416/Udhaar-Khata

How I Built It

Udhaar Khata is built from the ground up around open-weights AI models running locally on consumer hardware:

1. Information Extraction & Semantic Parsing (Google Gemma 2 2B)

We run Google's Gemma 2 2B (gemma2:2b) locally via Ollama with Metal GPU acceleration on Apple Silicon.
Gemma 2 is prompted to understand colloquial Indian financial terms (udhaar, jama, chukta, baaki, rokda, saman) and outputs strict structured JSON containing customer_name, amount, type, and note.
To guarantee 100% uptime and sub-millisecond responsiveness even under cold-start conditions, we paired Gemma 2 with a deterministic Indic-NLP bilingual rule engine.

2. Speech-to-Text (Open-Weight OpenAI Whisper)

Audio transcription runs through open-weight faster-whisper (ctranslate2 + int8 quantization) and the browser Web Speech API, allowing my father to speak naturally in Hindi (hi-IN) and Hinglish without typing a single character.

3. Polite Hindi Reminder Generator

In Indian business culture, asking customers to repay credit can feel awkward or socially sensitive.
Using Gemma 2 2B, the app drafts culturally nuanced payment reminders in three predefined tones (Polite, Friendly, Formal) as well as custom prompts (e.g. "Tell them distributor payment is due tomorrow"), and encodes them into instant WhatsApp click-to-chat links (https://wa.me/...).

4. Local Database & Privacy Architecture

All customer accounts, contact numbers, and transaction ledgers are stored in a local SQLite database on my father's laptop. Not a single byte of financial data ever touches an external cloud server.

Why Does Open Innovation Matter?

This project is a living example of where an open-weights approach worked infinitely better than closed commercial APIs:

Financial Sovereignty & Absolute Privacy: Small shopkeeper credit books contain confidential debt records, personal trust networks, and customer phone numbers. Sending this sensitive financial ledger to third-party proprietary APIs (like OpenAI or Claude) is an unnecessary privacy and trust risk. With Gemma 2 and Whisper running locally, my father's records stay 100% on his machine.
Works in Marketplaces with No Internet: Indian local markets and small towns frequently experience broadband cuts and spotty mobile reception. Closed APIs fail the moment the Wi-Fi drops. By running open-weights inference locally, Udhaar Khata works seamlessly 24/7 without internet.
Zero Operating Cost (Forever ₹0): A small neighbourhood shopkeeper operating on tight margins cannot pay metered token bills or $20/month SaaS subscription fees. Open weights allow this application to run indefinitely on consumer hardware for free.
Colloquial Linguistic Freedom: Open models gave us the flexibility to fine-tune system prompts for colloquial Hinglish dialects without hitting arbitrary API censorship or rigid guardrails around financial language.

My Agent Session

I built this project with pair programming and browser automation. You can explore the full session transcript here:

Building Udhaar-Khata: Voice AI Digital Ledger for Father using Local Gemma 2 and Whisper

You

I am participating in Hacktoberfest Weekend Challenge: Build for a Friend. Building Udhaar Khata for my father using local open-weight AI models (no closed APIs). He speaks in Hindi/Hinglish like 'Ramesh ne 200 udhaar liya' or 'Sunita ne 500 wapas kiye'.

Agent

Architecting Udhaar-Khata with 100% local open-weights AI: Google Gemma 2 (2B) running via Ollama on Apple Silicon for information extraction and polite reminder drafting, OpenAI Whisper for local speech-to-text, and a bilingual React interface with SQLite local storage.

Agent

Installed and configured Ollama with gemma2:2b locally. Verified local inference with sub-second Hindi/Hinglish transaction parsing and entity extraction.

Agent

Implemented polite payment reminder generator with multiple culturally respectful Hindi tones (Vinamra, Mitratapoorna, Vyavsayik) and 1-click WhatsApp deep integration. Tested end-to-end via browser automation.

Prize Categories

Best Use of Gemma (Featured Category)
Overall Hacktoberfest Winner