DEV Community: Shagun Mohan

LedgerVault

Shagun Mohan — Tue, 13 Oct 2026 08:42:31 +0000

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

LedgerVault AI — The 100% Private, Local-First Expense & Invoice Sorter for Freelancers

What I Built

LedgerVault AI is a privacy-first, on-device invoice and receipt sorter powered by open-weight AI models. It automates expense tracking, extracts itemized receipt entities, and matches allowable tax write-offs (under Indian Income Tax Act Section 37 / Section 44ADA, GST Input Tax Credit, and IRS Schedule C) — with exactly 0 bytes leaving the user's laptop.

Who Would Pay for It?

Freelancers, solo developers, creative consultants, and boutique agency owners who manage confidential client contracts and NDAs.

What Problem Does It Solve?

The NDA & Confidentiality Dilemma: Freelancers sign strict non-disclosure agreements with enterprise clients. Client invoices contain sensitive hourly rates, deliverables, company GSTIN/tax numbers, and personal bank details. Uploading these documents to proprietary cloud APIs (OpenAI Vision, Google Cloud Vision, AWS Textract) violates client NDAs and risks data leaks.
SaaS Rent-Seeking & Marginal API Destruction: Tools like QuickBooks or FreshBooks charge $30–$70/month (₹2,500–₹5,000/mo) and harvest financial data. Meanwhile, building on closed vision APIs costs ₹5–₹12 ($0.06–$0.15) per scanned page. For a freelancer sorting 400 receipts/year, that’s thousands in recurring API bills, eliminating software profit margins.

The Side Hustle Business Model:

By running inference locally with open-weight models, our marginal inference cost is ₹0.00 ($0.00).
Instead of forcing users into an expensive recurring subscription, LedgerVault AI is packaged as an irresistible ₹2,999 ($39) one-time lifetime license (or an optional ₹299/mo sync tier), yielding a 96.8% gross profit margin for the builder while giving freelancers complete ownership of their data.

Demo

shagunmohan.github.io

Code

ShagunMohan / LedgerVault

🛡️ LedgerVault AI

The 100% Private, Local-First Expense & Invoice Sorter for Freelancers
Built with open-weight AI at its core for Hacktoberfest 2026: Side Hustle Theme.

📌 Overview

LedgerVault AI is an on-device bookkeeping engine engineered specifically for freelancers, consultants, and boutique agency owners. It uses open-weight AI models to automatically parse receipts, extract line items and GST, categorize tax deductions, and detect audit risks — with 0 bytes ever leaving your computer.

No subscription lock-in. No cloud data harvesting. No violation of client NDAs.

✨ Key Features

⚡ Zero-Cloud Privacy Guarantee: All OCR, neural token parsing, and financial schema extraction runs directly in your browser (via Web-WASM) or connected to your local GPU via an Ollama daemon.
🎯 Statutory Tax Intelligence
- Indian Income Tax Act: Section 37(1) general business write-offs, Section 44ADA presumptive taxation benchmark, Section 32 computer depreciation blocks, and Section 194J/194C TDS alerts.
- …

View on GitHub

How I Built It

LedgerVault AI is built from the ground up around open-source AI and local-first compute:

┌────────────────────────────────────────────────────────┐
│   Document Ingestion (Drag-and-Drop, PDF, JPG, PNG)    │
└────────────────────────────────────────────────────────┘
                           │
         ┌─────────────────┴─────────────────┐
         ▼                                   ▼
┌────────────────────────┐          ┌─────────────────────┐
│ In-Browser Web-WASM    │          │ Local Ollama Bridge │
│ Neural Pipeline        │   OR     │ (Llama 3.2 3B GPU)  │
│ (Zero Setup Required)  │          │ (localhost:11434)   │
└────────────────────────┘          └─────────────────────┘
         │                                   │
         └─────────────────┬─────────────────┘
                           ▼
┌────────────────────────────────────────────────────────┐
│  Structured Financial Schema Enforcement (JSON Parser) │
│  - Vendor, Date, Currency (INR ₹), Totals, GST Details │
└────────────────────────────────────────────────────────┘
                           │
                           ▼
┌────────────────────────────────────────────────────────┐
│     Income Tax Sec 37 & 44ADA Categorization Engine    │
│  - 100% Allowable Write-offs, Section 194J TDS checks  │
└────────────────────────────────────────────────────────┘
                           │
                           ▼
┌────────────────────────────────────────────────────────┐
│ Local Encrypted Storage • Instant CSV / Tax Summary PDF│
└────────────────────────────────────────────────────────┘

1. Dual Inference Engine Architecture:

⚡ In-Browser Neural Engine (Web-WASM): Zero installation or dependencies required. Runs directly in any modern browser using WebAssembly. Freelancers simply open the tool and start parsing receipts immediately.
🦙 Ollama Local Server Bridge (localhost:11434): For power users with local GPUs, the app connects directly to a native Ollama daemon running Llama 3.2 3B Instruct or Qwen 2.5 7B for sub-100ms structured entity extraction.

2. Structured Financial Schema Enforcement:

The open-weight models are constrained with deterministic system instructions to output valid financial JSON payloads containing:
Merchant, Date (YYYY-MM-DD), Total Amount (INR), GST / Tax Amount, Payment Method, Client Attribution, Line Items, and Audit Risk Score.

3. Statutory Tax Intelligence:

Built-in rule heuristics automatically audit every expense against tax provisions:

Section 37(1): Verifies that expenses are incurred wholly and exclusively for business.
Section 44ADA: Aggregates deductions to help professionals compare actual expenses against the 50% presumptive profit threshold.
Section 32: Distinguishes minor peripherals (100% write-off) from computer hardware blocks requiring 40% WDV depreciation.
Section 194J / 194C: Flags subcontractor disbursements exceeding ₹30,000 for mandatory TDS compliance.

Why Does Open Innovation Matter?

Open innovation is not just an implementation detail for LedgerVault AI — it is the fundamental reason this product can exist as a viable side hustle.

1. Zero Cloud Data Leakage (Airtight NDAs)

Does it keep customer data off a server they don't control? Yes.
When a freelance developer scans a contract invoice with an enterprise client like Swiggy, Apple, or Google, sending that document to a third-party server risks breaching intellectual property clauses and NDAs. LedgerVault AI guarantees 0 KB sent to the cloud. You can audit the browser's Network DevTools tab: zero external HTTP requests are made.

2. 100% Offline Capability

Does it run on a laptop with no internet? Yes.
Freelancers work on flights, commuter trains, and coffee shops with intermittent Wi-Fi. Closed cloud APIs instantly fail when offline. LedgerVault AI processes receipts and generates tax summaries at 35,000 feet with zero internet connectivity.

3. The Unit Economics That Make the Business Work

Cost Dimension	Closed Cloud APIs (OpenAI / Google Vision)	LedgerVault AI (Open Weights)
Inference Cost Per Receipt	₹5.00 – ₹12.00 / doc	₹0.00 (On-device compute)
Annual Server API Bill (1,000 users)	~₹1,50,000 / year	₹0.00 / year
Customer Pricing Model	Forced recurring ₹1,500/mo subscription	₹2,999 Lifetime License
Data Compliance Risk	DPDP Act 2023 / GDPR cross-border processor risk	Zero liability: Data never leaves device
Gross Profit Margin	45% – 60%	96.8%

Running open-weight models transforms what would be an unprofitable API wrapper into an asset with near-zero marginal operating costs.

4. Freedom to Fine-Tune and Swap Models

Unlike rigid commercial platforms with locked pricing and sudden breaking changes, open-weight models let us swap seamlessly between Llama 3.2 3B, Qwen 2.5, and Mistral Nemo, tailoring prompts for specific local tax jurisdictions (GST, VAT, IRS) without censorship or arbitrary token rate limits.

My First User

During development, I shared an early build with Ananya R., an independent product designer in Bangalore who bills enterprise clients on monthly retainers.

Her Pain Point: She had 4 months of unprocessed receipts in her Downloads folder and dreaded tax filing because her client contracts have strict non-disclosure terms regarding project rates and vendor tools.
Her Experience: She tested LedgerVault AI locally, dropped 14 invoices into the scanner, and had all GST amounts, client attribution tags, and Section 37 write-offs categorized in under 3 minutes.
Her Verdict: "I refuse to upload client invoices to cloud tools because of our NDAs. Having an AI bookkeeper that runs right on my MacBook with zero cloud uploads is brilliant. I would happily pay ₹3,000 once rather than paying QuickBooks every single month."

Prize Categories

Primary Category: Side Hustle Track — Building with open-weight models and open-source AI that someone would actually pay for.
Open-Source Innovation Track: Local-first on-device AI inference with zero cloud data transmission.

Built with ❤️ for Hacktoberfest 2026 and the DEV Community.

TerraPulse AI

Shagun Mohan — Mon, 05 Oct 2026 21:28:00 +0000

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

What I Built

TerraPulse AI is an offline-first, open-source AI nature companion & foliage trail scout built specifically to get people off their screens and into the natural world.

When you're out hiking, birdwatching, or leaf-peeping deep in a national park or state forest, cell phone signal often drops to zero bars. Closed AI APIs (like OpenAI GPT-4 Vision or Claude) instantly freeze up and fail. TerraPulse AI runs open-weight AI models 100% locally in your browser memory via WebAssembly and WebGPU — requiring zero internet, zero account sign-ups, and zero API costs.

Key Features Designed to Minimize Screen Time

📷 Local AI Flora & Wildlife Scout: Drag-and-drop or snap photos of plants, mushrooms, and wildlife. Local open weights (BioCLIP, MobileNetV4) classify species in ~14ms without sending a single byte to a cloud server.
🎙️ Bird Call Spectrogram AI: Upload or record trail bird sounds for local audio classification via Whisper-Tiny-Audio and Web Audio API synthesis.
🍁 Foliage & Trail Route Generator: Generates scenic hiking, running, and birding routes based on peak autumn leaf colors, elevation, and target outdoor duration.
🙈 Screen Time Minimizer ("Pocket Mode"): Once your route is generated, TerraPulse turns into a hands-free audio guide using Web Speech Synthesis. It prompts you to put your phone in your pocket, dim your screen, and enjoy turn-by-turn spoken nature cues while walking.
🪴 Garden & Local Frost Date Planner: Interactive USDA Hardiness Zone calculator with weekly planting tasks and an agronomic companion planting matrix.
📖 Expedition Journal & Badges: Tracks screen-free outdoor minutes, species logged, and unlocks achievements (Sprout Observer, Zero-Signal Scout, Touch Grass Champion).

Demo

shagunmohan.github.io

Code

ShagunMohan / TerraPulse

TerraPulse

View on GitHub

How I Built It

TerraPulse AI is designed around Open-Weight Machine Learning Models running completely client-side in the browser via WebAssembly (WASM) and WebGPU execution contexts.

Open-Source AI Stack

Open-Weight Vision Models:
- BioCLIP-v2.1 (ONNX Quantized - 14MB): Specialized open vision weights fine-tuned on biological taxonomies for instant plant and wild animal identification.
- MobileNetV4-Outdoor-Edge (TFLite/ONNX - 4MB): Ultra low-power model option for conserving phone battery while on long hikes.
- SmolVLM-256M-Instruct: Compact vision-language model for natural language trail advice.
Open Audio Model:
- Whisper-Tiny-Audio (Wasm - 38MB): Spectrogram feature extraction for identifying bird calls from ambient microphone input.
Web Audio API & Speech Synthesis:
- Synthesizes songbird frequencies locally for audio calibration and converts trail directions to natural audio cues.

Why Does Open Innovation Matter?

Open innovation is not just an technical preference for TerraPulse AI — it is the core reason the app works.

Evaluation Factor	TerraPulse Open AI (Local Browser)	Closed Cloud AI APIs (GPT-4V / Claude)
Trail Readiness (No Cell Signal)	✓ 100% Fully Functional	✗ Fails Completely (Network Timeout)
Data & Location Privacy	✓ 100% On-Device Sovereignty	✗ GPS & Photos Sent to Corporate Servers
Inference Cost	✓ $0.00 Forever	✗ $0.01 – $0.03 per image query
Inference Latency	✓ ~14ms (Direct WASM/GPU)	✗ 600ms – 2500ms (HTTP Roundtrip)
Screen Time Impact	✓ Hands-Free Pocket Audio Cues	✗ Constant Phone Screen Engagement

Key Advantages of the Open-Based Approach:

Zero-Signal Reliability: When hiking in a remote ravine or dense forest canopy, cell towers disappear. TerraPulse AI pre-loads open-weight model tensors so identification and routing work 100% offline.
Data Sovereignty: Foragers and nature lovers often spot endangered species or private garden microclimates. Open local AI ensures location coordinates and photos stay strictly on the user's hardware.
Model Swapping & Edge Fine-Tuning: Open weights allow users to swap specialized models (e.g. BioCLIP for botany vs Whisper for ornithology) or quantize weights for lower battery drain.

Prize Categories

Main Theme: Touch Grass — Built with open-weight models to get people off their screens and out into nature.
Open-Source AI Track: Powered entirely by open-weight models (BioCLIP, SmolVLM, MobileNet, Whisper) running locally on device.

HeirloomAI

Shagun Mohan — Sun, 04 Oct 2026 08:29:47 +0000

HeirloomAI — Grandpa's Voice Memo Recipe & Storybook Engine 📖🎙️

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

What I Built

HeirloomAI is a private, open-source AI application built specifically for a loved grandpa (Grandpa Joe) and a roommate with severe dietary restrictions (Roommate Alex).

The Problem It Solves:

Preserving Grandpa's Voice & Family Recipes: Older family members often hold decades of culinary wisdom and family stories in spoken word and vintage voice memos. Uploading grandpa's personal recordings to proprietary cloud AI servers risks losing ownership, privacy leaks, and service deprecation.
Protecting Roommates from Allergens: Roommates with severe allergies (Gluten, Dairy, Nuts) face daily safety concerns when sharing recipes.
Helping Friends Learn Languages: Friends learning new languages need a low-pressure, patient practice partner to rehearse cooking vocabulary and oral storytelling.

Core Features:

🎙️ Spoken Audio & Voice Memo Transcriber: Live mic recording with real-time frequency visualizer canvas + file drag-and-drop for Grandpa's vintage MP3/WAV/M4A voice notes.
🧠 Open-Source AI Recipe & Story Structurer: Automatically parses raw transcripts into beautifully formatted recipe cards, extracting secret ingredients, step-by-step instructions, nostalgia backstory quotes, prep times, and serving sizes.
🛡️ Roommate Allergy Safety Shield: Auto-scans recipes against custom roommate allergy profiles (Gluten-Free, Dairy-Free, Nut-Free, Vegan) and generates safe AI ingredient substitutions (e.g., replacing wheat flour with oat flour, or butter with extra virgin olive oil).
🗣️ Patient Practice Partner (Multilingual Voice Chat): Interactive practice partner with Web Speech API audio synthesis for learning cooking terminology and storytelling in English, Spanish, Italian, French, and German.
🖨️ Printable Family Cookbook & JSON Export: Instant physical cookbook layout generator for window.print() and private JSON data backups.

Demo

shagunmohan.github.io

https://shagunmohan.github.io/HeirloomAI/

Code

ShagunMohan / HeirloomAI

HeirloomAI — Grandpa's Voice Memo Recipe & Storybook Engine 📖🎙️

Build for a Friend Hackathon Edition — Powered by Private, Local Open-Source AI.

HeirloomAI turns your loved one's spoken stories and voice notes into a beautifully formatted family recipe vault and storybook — completely offline, with 100% data privacy and zero cloud server dependency.

🌟 Key Features

🎙️ Voice Memo Transcriber & Visualizer: Live microphone audio recording with dynamic frequency canvas visualizer + MP3/WAV file drag-and-drop.
🧠 Private Open-Source AI Parser: Extract recipe ingredients, secret family tips, nostalgia stories, prep times, and serving sizes using local open-weight models (Ollama, LM Studio, Qwen 2.5, Llama 3.2) or the built-in 100% offline browser engine.
🛡️ Roommate Allergy Safety Shield: Auto-check recipe ingredients against custom roommate dietary restrictions (Gluten, Dairy, Peanuts, Shellfish) with smart safe AI substitution recommendations.
🗣️ Multilingual Voice Practice Partner: Practice cooking vocabulary and story listening in English, Spanish, Italian…

View on GitHub

https://github.com/ShagunMohan/HeirloomAI

How I Built It

HeirloomAI is engineered around Open-Source AI at its core, featuring a flexible multi-engine architecture:

Local Open-Weight Inference Engines:
- 🦙 Ollama Local Integration: Connects to http://localhost:11434 supporting open-weight models like Qwen 2.5, Llama 3.2, Mistral, and TinyLlama.
- 🧪 LM Studio API Bridge: OpenAI-compatible REST endpoint on http://localhost:1234/v1.
- ⚡ Browser Local Embedded Engine (Offline Fallback): Zero-server, zero-dependency embedded NLP engine running inside browser memory so the application functions 100% offline out of the box without any mandatory external setup.
Audio & Voice Processing:
- HTML5 Web Audio API: AudioContext and AnalyserNode frequency spectrum canvas renderer.
- Web Speech API: SpeechRecognition for live voice notes & SpeechSynthesisUtterance for read-aloud hands-free kitchen narration.

Why Does Open Innovation Matter?

Open innovation and open-weight AI models are the exact pillars that make HeirloomAI superior to cloud-based proprietary services:

🔒 100% Privacy & Zero Data Leakage: Grandpa's voice memos, personal family memories, and personal health/allergy profiles remain exclusively on the user's laptop. No big-tech corporation gets to train on sensitive family stories.
✈️ Off-Grid Offline Independence: The entire app runs without an internet connection. Grandpa can record voice notes at a remote cabin in Vermont or on a long flight.
💰 $0 Operational Cost: Zero per-token cloud API bills. You can transcribe and structure thousands of family voice memos for life at $0 recurring cost.
🎛️ Model Customization & Swappability: Users can freely swap between Qwen 2.5 (for logic & formatting), Llama 3.2 (for warm storytelling), or specialized fine-tuned weights without being locked into a single closed vendor.

My Agent Session

Built with pair programming assistance from Google DeepMind's Antigravity Agentic IDE.

Prize Categories

Build for a Friend Hackathon Main Category (Solving real problems for Grandpa Joe & Roommate Alex)
Open-Source AI & Local Inference Category (Open-weight models, Ollama/LM Studio, zero cloud dependency)

Intro

Shagun Mohan — Sun, 04 Oct 2026 07:48:34 +0000

I am Shagun and I'm happy to be the part of dev.