DEV Community: Abhinav Mehta

VaultInvoice AI: Zero-Knowledge Expense Auditor for Freelancers Built on Open-Weight Models

Abhinav Mehta — Tue, 13 Oct 2026 08:57:15 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

VaultInvoice AI is a zero-knowledge, local-first invoice and receipt auditor built for freelancers, independent contractors, and solo agency operators.

The Problem It Solves

Every tax season, freelancers lose hours deciphering messy folders of receipts, digital subscription PDFs, and subcontractor invoices. Existing expense tools like QuickBooks ($35/month), Dext ($30/month), and Expensify ($9/seat) suffer from two major flaws for solo operators:

Confidentiality & NDA Liabilities: Freelancers routinely handle sensitive client data—hourly rates, client addresses, banking details, and proprietary project scopes. Uploading these documents to cloud-based commercial AI models creates severe compliance risks.
Subscription Fatigue: Solo professionals resent paying $360+/year in recurring SaaS fees just to extract line items and sort tax deductions.

Who Would Pay For It?

Independent Developers & Designers: Who sign strict client Non-Disclosure Agreements (NDAs) and cannot upload vendor invoices containing client project details to public cloud endpoints.
Freelance Consultants & Accountants: Who want a one-time purchase ($49 license) or flat low-cost micro-SaaS desktop app that processes unlimited receipts with 100% data residency.
Digital Nomads & Field Workers: Who need to organize expenses offline while traveling or working in low-connectivity environments.

Demo

Local Preview: Running on http://localhost:5173/ with full real-time model streaming, token telemetry, and IRS Schedule C tax summaries.
Key Features:
- Ingestion of messy invoices and receipts (AWS cloud hosting, freelance subcontractor design bills, client dinners, equipment purchases).
- On-the-fly extraction of Vendor, Invoice #, Dates, Line Items, and Totals.
- Automatic classification into IRS Form 1040 Schedule C write-off buckets (Line 18 Office & Software, Line 11 Contract Labor, Line 24b 50% Meals, Line 13 Section 179 Equipment).
- Instant audit risk scoring flagging missing vendor tax IDs or unitemized cash payments.
- One-click export to Schedule C CSV, full audit JSON, and a copy-paste CPA briefing.
- Minimal Linear/Raycast-inspired design with inline item expansion and zero visual fluff.

Code

The complete source code is open source under the MIT License:
👉 Repository: https://github.com/Abh19avM/Hacktoberfest-Open-Source-AI-Challenge-dev-week2-

Tech Stack:

Core Engine: Open-weight LLMs via Ollama (llama3.1:8b, mistral:latest, gemma2)
Frontend: Vanilla CSS with sleek Linear/Raycast design tokens & Vite
UI Components & Visuals: Canvas Confetti, native responsive flex/grid
Proxy Layer: Vite local reverse proxy /api/ollama connecting directly to 127.0.0.1:11434 without CORS friction

How I Built It

VaultInvoice AI is built around open-weight models running on the user's local hardware:

flowchart TD
    A[Raw Invoice / Receipt Text or PDF] --> B[VaultInvoice UI]
    B --> C[Local Vite Proxy /api/ollama]
    C --> D[Local Ollama Engine 127.0.0.1:11434]
    D --> E[Open-Weight Model: Llama 3.1 8B / Mistral]
    E --> F[Structured JSON Parser]
    F --> G[IRS Schedule C Tax Classifier]
    G --> H[Audit Anomaly Detector]
    H --> I[Local Ledger & Export Suite]

Structured Open-Source Prompting

To guarantee deterministic financial output without hallucinations, we format the extraction prompt with strict JSON schema constraints and low sampling temperature (temperature: 0.1, top_p: 0.9):

{
  "vendor": "string",
  "invoiceNumber": "string",
  "date": "YYYY-MM-DD",
  "total": 0.00,
  "taxCategory": "Software & Cloud Hosting | Professional Services | Business Meals (50% Deductible) | Office Equipment (Sec 179)",
  "deductiblePercent": 100,
  "deductibleAmount": 0.00,
  "scheduleCLine": "Line 18 (Office Expenses)",
  "auditRisk": "Low | Medium | High",
  "auditReasons": ["string"]
}

The system streams tokens in real time directly to the browser console, displaying live speed metrics (tokens/sec and generation latency).

Why Does Open Innovation Matter?

Open innovation is not just an implementation detail for VaultInvoice AI—it is the exact reason this business model is viable where a closed API would fail:

Zero-Knowledge Data Privacy: Client billing records and confidential contracts remain strictly on the user's physical laptop. Zero packets ever leave the computer. If we built this with closed APIs (OpenAI or Anthropic), users would be forced to trust remote servers with their banking details and proprietary project names.
Sustainable Micro-SaaS Economics: Processing a 3-page itemized PDF or receipt batch via proprietary multimodal APIs costs $0.05 - $0.15 per document. For a freelancer processing 200 receipts a year, API costs erode margins. With open-weight models running on device, marginal inference cost is literally $0.00. We can sell a lifetime license for $49 and maintain a 98% gross margin.
True Offline Independence: Freelancers frequently travel or work in planes and coffee shops. Local models like Llama 3.1 run completely offline with no internet access.
No Vendor Lock-In or Sudden Deprecations: If a cloud provider doubles token prices or changes their terms of service, closed-API wrappers collapse. Open-weight models belong to the developer and the user forever.

My Agent Session

Here is the complete development session transcript showing how we architected, built, and tested VaultInvoice AI from start to finish:

Building VaultInvoice AI: Zero-Knowledge Expense Auditor for Freelancers

You

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My Agent Session

My First User

Prize Categories

<USER_SETTINGS_CHANGE>
The user changed setting Model Selection from None to Gemini 3.8 Flash (High). No need to comment on this change if the user doesn't ask about it. If reporting what model you are, please use a human readable name instead of the exact string.
</USER_SETTINGS_CHANGE>

Agent

▸ list_dir {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ read_url_content {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ grep_search {}

Input

{
}

Output

Completed successfully

Agent

▸ grep_search {}

Input

{
}

Output

Completed successfully

Agent

▸ grep_search {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ grep_search {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ read_url_content {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ manage_task {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ read_url_content {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

You

bro what are you doing ?

Agent

I was pulling the official rules, rubrics, and sponsor categories directly from the DEV API for the Hacktoberfest Week 2: Side Hustle Challenge so we know exactly what the judges are looking for!

Here is the quick breakdown:

The Theme: Side Hustle — Build a tool powered by open-source AI / open-weight models that solves a real problem someone would actually pay for (freelancers, small shop owners, niche creators, etc.).
The "Why Open" Requirement: Your submission must clearly explain why open-source AI makes this viable over closed APIs (e.g., zero API costs = sustainable micro-SaaS margins, 100% local privacy for sensitive customer data, runs completely offline, model customization).
Deadline: October 18, 2026 (1 overall winner + 17 partner category winners).
Bonus: Testing it with your first real user for feedback!

Where do you want to start?

Do you already have a project in mind, or would you like to pick from a few high-impact concepts:

Local Receipt & Invoice Privacy Sorter for Freelancers: Runs local OCR/vision + an open-weight model (like Gemma/Llama via Ollama/Transformers) to categorize expenses, extract line items, and export tax summaries—client financial data never leaves their laptop.
E-Commerce Listing & Pricing Copilot for Sellers (Etsy/eBay/Shopify): Turns photos and bullet points into platform-tailored, SEO-optimized descriptions and tags with zero recurring per-token API fees.
Hyper-Local Restaurant Menu & Allergen Translator: Fine-tuned/prompted on regional culinary dish descriptions, ingredients, and allergen warnings that generic commercial translators butcher.
Your own idea: Tell me what you're thinking, and we can architect the tech stack, build the app, and draft the DEV submission post.

You

Challenge Prompt
Side Hustle
Build something with open-source AI at its core.
That can mean running an open-weight model, building on an open-source agent harness or framework, running inference locally, or all three. Whatever you pick, the open pieces should be what makes your project work.
In your post, tell us why open innovation matters for what you built. Does it run on a laptop with no internet? Keep someone's data off a server they don't control? Let you fine-tune, swap models, or change how your agent behaves? Cost nothing to run? Tell us where your open-based approach worked better than a closed one.
This Week's Theme: Side Hustle
Build something with open-weight models or open-source AI that someone would actually pay for.
That can mean running an open-weight model, building on an open-source agent harness or framework, running inference locally, or all three. Whatever you pick, the open pieces should be what makes your project work.
Think small, specific, and useful. The best builds here solve a real problem for a real person, whether that's a freelancer, a small business owner, a niche hobby community, or your friend who keeps complaining about the same tedious task. A few ideas to get you going:
An invoice and receipt sorter for freelancers that runs locally, so client data never leaves their laptop
A listing writer for Etsy or eBay sellers that turns a few product photos into a polished description
A menu translator for a local restaurant that handles their specific dishes better than a generic tool
In your post, tell us why open innovation matters for what you built. Does running your own model keep costs low enough that the business actually works? Keep your customers' data private? Let you fine-tune for a niche that a general-purpose API handles poorly? Avoid getting locked into someone else's pricing? Tell us where your open-based approach worked better than a closed one.
Bonus points if you land your first user. Get it in front of someone, let them use it, and tell us how it went
<truncated 576 bytes>
3%23%20Demo%0A%3C%21--%20Share%20a%20deployed%20link%20or%20a%20video%20demo.%20--%3E%0A%0A%23%23%20Code%0A%3C%21--%20Show%20us%20the%20code%21%20%20You%20can%20embed%20a%20GitHub%20repo%20directly%20into%20your%20post.%20--%3E%0A%0A%23%23%20How%20I%20Built%20It%0A%3C%21--%20Which%20open-source%20AI%20did%20you%20use%20%28open-weight%20models%2C%20agent%20harnesses%2C%20frameworks%2C%20local%20inference%29%2C%20and%20how%20is%20your%20project%20built%20around%20it%3F%20--%3E%0A%0A%23%23%20Why%20Does%20Open%20Innovation%20Matter%3F%0A%3C%21--%20Why%20does%20open%20innovation%20matter%20for%20what%20you%20built%3F%20%20What%20did%20it%20make%20possible%20that%20a%20closed%20API%20wouldn%27t%3F%20--%3E%0A%0A%23%23%20My%20Agent%20Session%0A%3C%21--%20Optional%2C%20but%20judges%20love%20it.%20%20Save%20your%20session%20with%20DevRelay%20and%20embed%20it%20with%20the%20agent_session%20tag%20%28see%20the%20challenge%20page%29%2C%20or%20link%20to%20it.%20--%3E%0A%0A%23%23%20My%20First%20User%0A%3C%21--%20Optional%2C%20but%20this%20is%20where%20bonus%20points%20live.%20%20Did%20someone%20use%20it%3F%20%20Who%20were%20they%2C%20what%20did%20they%20say%2C%20and%20did%20they%20pay%3F%20%20If%20not%2C%20remove%20this%20section.%20--%3E%0A%0A%23%23%20Prize%20Categories%0A%3C%21--%20Which%20partner%20categories%20are%20you%20entering%3F%20%20List%20every%20one%20that%20applies%2C%20or%20remove%20this%20section.%20--%3E%0A%0A%3C%21--%20Team%20Submissions%3A%20Please%20pick%20one%20member%20to%20publish%20the%20submission%20and%20credit%20teammates%20by%20listing%20their%20DEV%20usernames%20directly%20in%20the%20body%20of%20the%20post.%20--%3E%0A%0A%3C%21--%20Thanks%20for%20participating%21%20--%3E%0A)Show your work. We'd love to see how you built it. Save your agent session with DevRelay and embed it in your post, or link to it. It's optional, but it helps the judges understand your process.

Agent

▸ ask_question {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ list_dir {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ manage_task {}

Input

{
}

Output

Completed successfully

Agent

▸ manage_task {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ list_dir {}

Input

{
}

Output

Completed successfully

Agent

▸ list_dir {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ manage_task {}

Input

{
}

Output

Completed successfully

Agent

▸ manage_task {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ manage_task {}

Input

{
}

Output

Completed successfully

Agent

▸ browser_subagent {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

🛡️ VaultInvoice AI: Zero-Knowledge Local Expense Auditor

We have built and launched VaultInvoice AI—a local-first, privacy-guaranteed invoice and receipt auditor designed for freelancers, independent contractors, and solo businesses.

The application is currently running live on your machine at http://localhost:5173/!

🌟 What Makes This Stand Out for the Challenge

Open-Source AI at its Core:
- Connects directly to your local Ollama daemon (http://[REDACTED]:11434) and auto-detects your installed open-weight models (llama3.1:8b, mistral:latest, deepseek-coder:6.7b, gemma4:latest).
- Includes an intelligent deterministic local engine fallback so it functions seamlessly in any environment.
- Streams tokens in real time directly to the browser console with live throughput telemetry (tokens/sec and latency).
A Real Solution to a Real Problem ("Side Hustle" Theme):
- Data Confidentiality & NDAs: Freelancers routinely handle sensitive client data (hourly rates, client addresses, and proprietary project scopes). Commercial cloud APIs violate client NDAs; VaultInvoice guarantees 0 bytes leave the machine.
- Micro-SaaS Margins: Cloud LLM APIs charge $0.05–$0.15 per complex receipt parse, eroding solo operator margins. Running on local open weights means marginal API cost is $0.00.
Smart IRS Schedule C Mapping & Audit Radar:
- Automatically categorizes line items into IRS write-off buckets:
  - Line 18: Office & Cloud Software (100% deductible)
  - Line 11: Subcontractor Labor (100% deductible + Form 1099-NEC threshold alert)
  - Line 24b: Business Meals (50% statutory limitation under IRC § 274)
  - Line 13: Equipment Depreciation & Section 179
- Anomaly detector flags suspicious cash transactions or missing vendor tax IDs.
Export Suite:
- Schedule C Itemized CSV (ready for Excel/TurboTax/QuickBooks)
- Full Audit JSON
- **Copy-to-Clipb

You

nah bro this seems too much ai generated i want sleat minimal design of the UI and very much self explnatory :)

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ browser_subagent {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

I’ve completely stripped out the noisy badges, flashy neon glows, and cluttered modals in favor of a sleek, minimal Linear / Raycast-style interface.

What Changed:

Calm, High-Contrast Palette: Pure neutral dark tones (#09090b), clean 1px borders (#27272a), and crisp typography with no unnecessary glowing gradients.
Self-Explanatory & Direct:
- Top: Minimal brand mark, quiet local model indicator (● llama3.1:8b), Export CSV, and Clear.
- 3 Calm Numbers: Total Incurred ($), Tax Write-Offs ($), and Privacy Status (100% Local).
- Left Panel (Input): Clean drag-and-drop zone + 1-click sample links (AWS, Contractor, Dinner, Monitor) + clean textarea + Analyze Receipt button.
- Right Panel (Ledger): Minimal table rows showing Vendor, Date, Category tag (Software · 100%, Meals · 50%), Amount, and Deductible write-off.
Inline Expansion (No Clunky Popups): Clicking any row smoothly reveals the itemized breakdown, IRS Schedule C line, and compliance notes right in place.

Refresh http://localhost:5173/ to see the clean new design!

You

so how would i submit this :)
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My Agent Session

My First User

Prize Categories

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

Here is the step-by-step guide to submitting your project for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle:

Step 1: Push the Code to a Public GitHub Repository

In your terminal or PowerShell (in d:\Hacktoberr\dev-week-2):

# 1. Create a new empty repository on GitHub (e.g. named "vault-invoice")
# 2. Link your local repo to GitHub and push:
git remote add origin https://github.com/<YOUR-GITHUB-USERNAME>/vault-invoice.git
git branch -M main
git push -u origin main

Step 2: Open the DEV Challenge Submission Editor

Click this official prefilled challenge submission link:

👉 Open DEV Challenge Editor

Ensure the tags include:
devchallenge, hf26challenge

Step 3: Copy & Paste Your Submission Article

I have pre-written the complete submission article adhering directly to the judging rubric in DEV_SUBMISSION.md.

Here is the exact text to copy into your post:

---
title: VaultInvoice AI: Zero-Knowledge Expense Auditor for Freelancers Built on Open-Weight Models
published: false
tags: devchallenge, hf26challenge, ai, opensource
---

*This is a submission for the [Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle](https://dev.to/challenges/hacktoberfest-week2-2026-10-12)*

## What I Built

**VaultInvoice AI** is a zero-knowledge, local-first invoice and receipt auditor built for freelancers, independent contractors, and solo agency operators.

### The Problem It Solv

You

Abh19avM/Hacktoberfest-Open-Source-AI-Challenge-dev-week2-

push in this repo :) and give me draft for dev post also it mentioned some
"Show your work. We'd love to see how you built it. Save your agent session with DevRelay and embed it in your post, or link to it. It's optional, but it helps the judges understand your process." how can we do that too i ig i have devrelay

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ replace_file_content {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ view_file {}

Input

{
}

Output

Completed successfully

Agent

▸ replace_file_content {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

Agent

▸ write_to_file {}

Input

{
}

Output

Completed successfully

Agent

▸ run_command {}

Input

$ {}

Output

Completed successfully

(Direct link: https://dev.to/agent_sessions/building-vaultinvoice-ai-zero-knowledge-expense-auditor-for-freelancers-limi5e)

Prize Categories

Overall Open-Source AI Challenge Winner

CanopyRoute Pune — Offline Open-Source AI Trail Audio Runner

Abhinav Mehta — Sun, 11 Oct 2026 08:04:05 +0000

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

What I Built

Most modern outdoor and trail running apps create a paradoxical problem: they encourage us to get out into nature, but keep our eyes glued to bright phone screens checking turn-by-turn prompts, heart rate charts, and elevation contours.

I built CanopyRoute Pune to turn this dynamic on its head: make the screen the shortest part of the run, put your phone in your pocket, and run by sound.

Designed specifically for the tekdis (hills), banyan groves, and mountain crests of Pune, Maharashtra, CanopyRoute is an offline-first scenic trail route planner and auditory hill running companion powered by local open-weight AI.

Core Features

Touch Grass Pocket Run Mode: Select your trail, tap "Start Pocket Run", and slide the phone into your running shorts or vest. The screen switches to an ultra-low-power OLED dark display while in-browser speech synthesis and haptics deliver sensory grounding cues (e.g., "Notice the earthy petrichor scent of damp red murrum soil warming up under the morning sun") as you pass GPS milestones.
Natural Language Discovery: Instead of rigid dropdown filters, runners describe their desired run in natural language ("Quiet 5k morning run on Vetal Tekdi with peacocks, soft red soil, and banyan trees"). An on-device open-weight embedding model ranks trails by terrain, canopy density, and flora.
Accurate Regional Cartography: Powered by Leaflet with high-resolution ESRI Topographic & Satellite tiles strictly bounded to Pune and the Western Ghats (minZoom: 10, maxBounds set to Sahyadri coordinates), ensuring authentic local hill contours without border clipping issues.
Interactive Cursor Companion: A smooth, physics-based flying vector companion follows your cursor across the screen with natural banking and directional rotation. You can switch between the Blue Mormon Butterfly (Maharashtra's state butterfly with 3D flapping wings) and the Hill Honeybee with a sparkling pollen dust trail.
Curated Pune Sahyadri Trails:
- Vetal Tekdi & ARAI Forest Ridge (5.8 km • Highest crest in Pune, red murrum soil, peacocks, wild hares)
- Taljai Hills & Peacock Reserve (4.6 km • Protected forest reserve with shaded banyan groves and resident dancing peacocks)
- Sinhagad Fort Mountain Trail (8.2 km • Historic basalt steps, cool valley fog, Karvi flower blossoms, Khadakwasla reservoir views)
- Pashan Lake Bamboo & Wetland Sanctuary (4.2 km • Golden bamboo groves, lotus water lilies, and kingfishers)
- Baner Hill & Windmill Plateau Ridge (5.2 km • Panoramic 360-degree ridge run overlooking the Baner-Pashan valley)
Pune Botanical & Wildlife Field Guide: An offline field guide to native Western Ghats trees (Banyan aerial prop roots, Gulmohar scarlet blooms, Karvi mass-flowering shrubs, Champa) and wildlife (Indian Peacock, Blue Mormon Butterfly, Black-naped Hare, Hill Honeybee).

Demo

Live Deployed App: https://abh19avm.github.io/Touch-Grass-hacktober-week1-/
Repository: https://github.com/Abh19avM/Touch-Grass-hacktober-week1-

(Tip: Open the live link on mobile or desktop, describe a trail in plain English, click "Pocket Run Mode", and listen to the procedural nature soundscape and live landmark voice cues!)

Code

Abh19avM / Touch-Grass-hacktober-week1-

🦚 CanopyRoute Pune — सह्याद्री Trail Running & Nature Audio Companion

Built for Hacktoberfest 2026 Open-Source AI Challenge: Week 1 ("Touch Grass")
Theme: Build something with open-source AI at its core that gets people off the screen and into the world.
Live Site: https://abh19avm.github.io/Touch-Grass-hacktober-week1-/
Repository: https://github.com/Abh19avM/Touch-Grass-hacktober-week1-

🌲 What is CanopyRoute Pune?

CanopyRoute Pune is an offline-first scenic trail route planner and auditory hill running companion powered by local open-weight AI (@huggingface/transformers), built specifically for the tekdis, banyan groves, and mountain crests of Pune, Maharashtra.

Modern fitness and mapping apps keep runners glued to glowing screens, constantly staring at digital turn-by-turn prompts, elevation contours, and pace alerts. CanopyRoute is built with the opposite philosophy: make the screen the shortest part of the experience, put the phone in your pocket, and run by sound.

✨ Core Features

Touch Grass Pocket Run Mode
- Select your trail, tap "Start Pocket…

View on GitHub

How I Built It

CanopyRoute is engineered to run 100% locally on the runner's device with zero external server dependencies:

Local Open-Weight AI Inference (@huggingface/transformers):
- Runs Xenova/all-MiniLM-L6-v2 locally in the browser using WebAssembly / WebGPU via ONNX Runtime Web.
- When a runner types a sensory query like "Lush post-monsoon banyan grove with peacocks and soft dirt", the model computes 384-dimensional vector embeddings on-device and calculates cosine similarity against curated trail document vectors in milliseconds.
- Includes a deterministic, zero-network edge vectorizer fallback for instantaneous query resolution on resource-constrained mobile browsers.
Procedural Auditory Coaching Engine:
- Instead of streaming pre-recorded audio files over cellular networks, CanopyRoute uses the Web Audio API to procedurally synthesize ambient mountain breezes and wind noise using pink noise buffers and biquad resonant filters.
- Combined with the Web Speech Synthesis API, it articulates mindful landmark cues, heart-rate pacing reminders, and botanical observations directly through your headphones while the screen is locked in your pocket.
Cartography & Leaflet Integration:
- Built with React-Leaflet and Leaflet, centered on Pune ([18.5240, 73.8180]).
- Regional constraints (maxBounds: [[17.80, 73.20], [19.20, 74.40]]) ensure runners explore authentic Pune hill tracks with high-resolution ESRI Topographic and Satellite tile layers.
Vector Illustrations & CSS 3D Animations:
- Built completely without generic emojis. Features custom SVG vectors with 3D transform animations (rotateY wing flapping on the Blue Mormon butterfly, high-frequency buzz on the honeybee, and spring-smoothed mouse tracking).

Why Does Open Innovation Matter?

Mountain Trails Have Zero Cell Service:
When running through the deep forested valleys of Sinhagad Fort or behind the ridges of Taljai, cloud-based proprietary APIs (like OpenAI or Anthropic) fail instantly due to lack of connectivity. Open-weight models downloaded to browser cache execute seamlessly offline on mountain summits without a single bar of cell reception.
Uncompromising Biometric & Location Privacy:
Running routes, GPS coordinates, pacing, heart rate, and schedule patterns are sensitive personal telemetry. With closed cloud APIs, this data is sent over the wire to remote servers. With open-source on-device AI, zero bytes of biometric or location data ever leave the runner's phone.
Zero Cost for Grassroots Running Clubs:
Non-profit trail conservancies and community running groups in cities like Pune can host and share this tool indefinitely with zero recurring API subscription fees or token bills.
Freedom to Adapt for Local Ecosystems:
Open innovation allows developers anywhere in the world to tailor environmental models to their local bioregion—from the red murrum soil and banyan trees of Maharashtra to alpine pine forests across the globe.

Prize Categories

Hacktoberfest Open-Source AI Challenge: Week 1 (Touch Grass)
Best Offline / Edge AI Experience
Open Innovation & Community Impact