Harshit Agarwal

Posted on Oct 15

DukaanVikaas: Offline Kirana AI Billing

#devchallenge #hf26challenge

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission 💸

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

What I Built

DukaanVikaas (दुकान विकास) is a 100% browser-native, offline-first inventory and point-of-sale (POS) billing application crafted specifically for India's 12+ million neighborhood Kirana stores and micro-retailers.

It enables shopkeepers to record sales using conversational, regional language commands (such as "5 packet Maggi bike", "3 Thums Up sell thai", "2 packet biscuit bech diye", or compound sentences like "Aaj 5 packet Maggi aur 3 cold drink bike"). It tracks inventory levels atomically, warns against stockouts, generates dynamic revenue analytics, supports instant one-click bill reversals with compensating return audit logs, and allows offline versioned JSON backups—all running with zero latency directly within the merchant's browser.

Who Would Pay For It & What Problem Does It Solve?

The Problem: Small Indian Kirana stores operate on razor-thin margins in fast-paced retail environments, often in semi-urban or rural areas plagued by erratic power and intermittent internet. Traditional enterprise billing software (like Tally, Marg, or Vyapar) is expensive, hardware-dependent, clunky, and overwhelmingly English-first. Conversely, cloud-only SaaS apps crash when the internet drops in the middle of a rush-hour customer queue.
The Solution: DukaanVikaas requires zero server backend, zero cloud databases, zero API keys, and zero subscription fees. It runs on any inexpensive Android smartphone, tablet, or desktop browser.
Who Pays / Economic Value: Small retailers, FMCG distributors, and micro-merchants save thousands of rupees in monthly software subscriptions, hardware POS rentals, and lost revenue from network downtime.

Demo

🌐 Live Deployed Application: https://dukaan-vikaas.vercel.app/
💻 GitHub Repository: https://github.com/Harshitagarwal113/DukaanVikaas

Key Workflows to Explore in the Demo:

Natural Language Assistant: Tap "AI Bill Entry" on the dashboard. Click sample prompts like "5 packet Maggi bike" or "3 Thums Up sell thai", parse the command, observe automatic price derivation and stock check, and explicitly confirm the sale.
Ambiguity Safety Guard: Try typing an ambiguous command like "2 cold drink bike". Notice that the engine never silently guesses—it identifies candidates (Thums Up, etc.), prompts for exact confirmation, and disables checkout until verified.
Instant Multilingual Switching: Switch between हिन्दी (Hindi), ગુજરાતી (Gujarati), Hinglish, and English via the top header selector with zero page reload.
Stock Integrity & Sale Reversal: Navigate to Sales History, inspect any completed receipt, and click Reverse Sale. Watch the item atomically return to inventory stock, an audit trail movement generated with reason 'return', and today's revenue recalculate dynamically.
JSON Backup & Restore: Open Settings, download dukaan-vikaas-backup.json, inspect the schema validation, or preview before restoring.

Code

Harshitagarwal113 / DukaanVikaas

DukaanVikaas — Browser-Native Kirana Store Inventory & Sales Manager with Offline Simulated NLP

DukaanVikaas (दुकान विकास) 🛒🌾

A minimal, high-performance, browser-native Kirana inventory and point-of-sale billing web application designed specifically for Indian neighborhood retail stores.
Built with an offline-first architecture powered completely by client-side IndexedDB, deterministic simulated natural-language transaction parsing, and browser-native JSON backups.

🌐 Live Application: https://dukaan-vikaas.vercel.app/
📦 GitHub Repository: https://github.com/Harshitagarwal113/DukaanVikaas

📖 Table of Contents

💡 Core Philosophy

Traditional retail SaaS platforms require continuous high-speed internet, monthly cloud subscription fees, complex authentication, and external databases that fail when network connectivity drops in rural or semi-urban areas.

DukaanVikaas is…

View on GitHub

The entire codebase is open-source, strictly typed, and structured into clean, modular layers:

lib/ai/simulatedParser.ts: 100% offline deterministic rule-based natural language parser with multilingual tokenizers, phonetic alias matching, and candidate suggestion guards.
lib/storage/db.ts: Client-side IndexedDB multi-store architecture with atomic transactional integrity and signed audit movement logs.
lib/validation/index.ts: Pre-import backup validator preventing corrupt schemas, negative pricing, and duplicate IDs.
lib/i18n/LanguageContext.tsx: Persistent vernacular state engine with memoized synchronization across local storage and IndexedDB.
tests/runUnitTests.ts: 45-point automated test suite validating the NLP parser, business math, and schema integrity.

How I Built It

Rather than relying on heavy, black-box cloud LLM APIs—which charge per token, introduce latency, require internet access, and risk catastrophic hallucinations in financial accounting—DukaanVikaas was designed around an open-source, client-side deterministic NLP engine and local-first architecture:

1. Local Deterministic NLP Engine

Conversational Tokenizer: Cleans regional conversational prefixes (aaj, bhaiya, sethji, sun, please) and verb suffixes (bike, bika, bech diye, sell thai, vechya, aapi).
Multilingual Numeral Normalization: Maps written number words in Hindi, Gujarati, Devanagari, and English (teen, paanch, tran, દો, પાંચ, five) and fractional terms (aadha, half) to numbers.
Compound Sentence Parser: Splits compound commands via regional conjunctions (aur, and, ane, तथा, commas) into separate transaction line items.
Fuzzy Catalog & Alias Scoring: Matches items against product names, categories, and vernacular aliases (मैगी, આટા, Desh Ka Namak).
Zero-Guess Ambiguity Guard: If an input is ambiguous (e.g. general terms like "biscuit" or "cold drink"), the engine classifies the item as needs_clarification, presents quick-pick buttons, and blocks checkout until the merchant explicitly confirms the item.

2. Browser-Native Storage & Tech Stack

Next.js 16 (App Router + Turbopack) & React 19: Ultra-fast client rendering with sub-second hot reload and zero build bloat.
Tailwind CSS v4: Minimalist Forest Green retail SaaS aesthetics (#065f46), high-contrast charcoal typography, neutral borders, and touch-optimized buttons ($\ge 44\text{px}$).
Browser-Native IndexedDB: 5 transactional object stores (products, sales, inventoryMovements, settings, metadata) providing offline durability.
Native JSON Blob Backup: Browser-triggered downloads without file system vulnerabilities or arbitrary disk access claims.

Why Does Open Innovation Matter?

Open innovation is essential for the future of everyday computing, especially in developing economies:

Digital Sovereignty & Absolute Privacy: A shopkeeper's daily revenue, wholesale margins, customer credit (khata) records, and supplier relationships are their most confidential business data. Closed cloud platforms harvest and monetize this data. With DukaanVikaas's open client-native architecture, 100% of data stays inside the merchant's physical browser—no server logs, no tracking, and zero leakage.
Resilience Against the "Always-On" Myth: Big tech software often assumes continuous gigabit fiber and reliable power. Real-world retail in emerging markets operates under brownouts and cellular dead zones. Open client-side architectures ensure technology adapts to human realities, rather than demanding costly infrastructure upgrades.
Zero Financial Friction & True Democratization: Closed AI services lock users into recurring token billing and subscription tiers. Open, browser-executable logic costs $0.00 to run. A micro-retailer in an Indian village has access to the exact same cutting-edge natural language billing efficiency as a luxury supermarket in a metropolitan hub.
Predictability Over Hallucination: Closed LLMs are probabilistic—they can miscalculate currency, hallucinate prices, or invent inventory items. In point-of-sale accounting, an error of ₹10 is a breach of customer trust. Open deterministic algorithms guarantee transparent, auditable math every single time.

My Agent Session

This application was developed as a solo engineer using paired agentic software engineering with DeepMind's Antigravity IDE and DevRelay:

Test Automation: Used subagents to perform end-to-end browser QA across mobile viewports (390x844), tablet screens, and desktop layouts.
Continuous Validation: Executed automated unit tests (45/45 passed), strict ESLint checks, and production builds with zero warnings.
Session Record: The full autonomous workflow, code edits, browser verification runs, and Git commit trails were orchestrated and verified end-to-end locally.

My First User

During development, DukaanVikaas was demonstrated to Ramesh-ji, an independent Kirana store owner operating a neighborhood provision shop:

The Reaction: "Ye to badiya hai! Padhne-likhne ki zarurat nahi, bas bolo '5 packet Maggi bike' aur bill ban gaya!" ("This is great! No need to search through complicated menus—just say '5 packet Maggi bike' and the bill is generated!").
Key Insight: He particularly appreciated that the app worked in Hindi and Gujarati without lagging, and that he could back up his complete shop data to a data.json file on his phone whenever he wanted, giving him complete peace of mind that no cloud company could ever lock him out of his records.

Prize Categories

Participation: Solo Developer (Harshit Agarwal / @Harshitagarwal113)

Open Source AI / Local-First Innovation: For building an offline, deterministic, zero-cloud natural language retail interface running 100% client-side in the browser.
Tech for Good & Social Impact: For empowering unserved micro-retailers (Kirana stores) with free, localized, accessible billing software that requires no expensive hardware or subscriptions.
Web Applications / Edge Computing: For showcasing Next.js 16, React 19, and IndexedDB with zero server dependency and sub-second performance.

DEV Community