DEV Community: SUMEDHA GHOSH

Two Cultures, One Night: I Built an Altar for the People We Miss

SUMEDHA GHOSH — Sat, 31 Oct 2026 13:28:50 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

What I Built

I'm a first-year BTech student, and until this month I could only write Java loops and basic HTML/CSS. So I took the "Ghost in the Machine" theme and asked myself: who are the ghosts I actually care about?

The answer was our ancestors. Mexico's Día de los Muertos and India's Bhoot Chaturdashi are thousands of kilometres apart, but both are built on the same idea: on one night, the people we lost are welcomed back with light. Mexican families build ofrendas with marigolds and candles. On Bhoot Chaturdashi, 14 diyas are lit for 14 ancestors.

Smriti & Ofrenda (smriti means "memory" in Hindi/Bengali) lets you:

Switch between Día de los Muertos and Bhoot Chaturdashi mode
Tell the app about someone you miss: their name, favourite food, favourite song, and a memory
Get a short, warm remembrance written by an open-weight AI model
Build a virtual altar, or light 14 diyas, one for each ancestor
Watch ghost emojis 👻 drift across the screen while you do it

Demo

👉 Try Smriti & Ofrenda here

![Día de los Muertos mode with marigolds, candles and the AI remembrance](

)

![Bhoot Chaturdashi mode with 14 diyas lit](

)

A few things to try:

Enter a name, a favourite food and a song, then press Generate Remembrance
Switch to Bhoot Chaturdashi and light all 14 diyas. The last one unlocks a message.
Look at the background and spot the ghosts. 👻

Code

GitHub repo link

How I Built It

I designed the screens in Figma, then turned them into a working app with a text-to-code tool, writing a detailed prompt that described every feature. I am a beginner, so that was my way into building something real.

The AI part is [Llama-3.2-1B-Instruct / Gemma, whichever your app uses], an open-weight model that runs through WebLLM directly in the browser. There is no backend, no API key and no paid service. When you press Generate Remembrance, the model writes the message on your own device, and the text changes depending on which tradition you picked.

Why Does Open Innovation Matter?

This is the part I care about most.

A remembrance app asks people for the most private things they have: a grandmother's name, her favourite sweet, a story only the family knows. I did not want that going to a server I don't control.

Because the model is open-weight and runs locally:

Privacy: the memories never leave your device
Cost: it is free to run, so there is nothing to pay and no account to make
Permanence: an altar for the people we lost should not disappear because a company changes its pricing or shuts down an API
Offline: after the first model download, it can work without internet [verify this before keeping it]

A closed API could have written nicer sentences. But it would have made me choose between convenience and my family's privacy. Open models meant I didn't have to choose.

The Spirit of It

[WRITE 3-4 REAL SENTENCES HERE. Who did you think of while testing? Which person's name did you type first, and what did the app write back? Did a particular line from the AI make you pause? Keep it honest and specific. This is the part that will make your post stand out.]

Prize Categories

[Check the challenge page. List a partner category only if you really used that technology, otherwise delete this whole section.]

Built by [@sumedha_ghosh_302008]. Happy Halloween, Día de los Muertos, and Bhoot Chaturdashi. 🕯️

ClutterSloth: Sorting 5 messy files took me 12 minutes of procrastination, so now a local browser loop does it while I nap

SUMEDHA GHOSH — Sun, 25 Oct 2026 15:18:18 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

ClutterSloth is a zero-dependency, local-first file and inbox tidier that organizes chaotic filenames (invoice_v1(1).pdf, Q4_meeting_notes_FINAL-final.docx, IMG_8842 copy.jpg) into clean, structured categories without sending a single byte over the network.
• Live Deployed Prototype: https://pug-tape-64702791.figma.site/
• The Core Rule: 0 bytes leave this browser. The sorting loop runs client-side, proving that you do not need heavy cloud infrastructure or expensive proprietary LLM APIs to automate everyday digital annoyance.

Demo

A walkthrough of the live interface: inputting messy filenames into the left control panel, toggling Local RAM Mode (Offline), and executing the client-side sort loop to populate structured folder outputs.
• ClutterSloth Live Prototype

Code

You can inspect and fork the interface architecture directly via the Figma Make Prototype Source or embed the client-side classification loop:
// ClutterSloth Client-Side Sort Loop
function runClutterSloth(rawInput) {
const lines = rawInput.split('\n').map(l => l.trim()).filter(Boolean);
return lines.map(filename => {
const lower = filename.toLowerCase();
if (lower.includes('invoice') || lower.includes('pdf')) {
return { original: filename, folder: '📁 /Finance_Receipts', action: 'Standardized nomenclature' };
} else if (lower.includes('note') || lower.includes('doc')) {
return { original: filename, folder: '📝 /Work_Notes', action: 'Tagged [Offline-Summary]' };
} else if (lower.includes('img') || lower.includes('jpg') || lower.includes('mov')) {
return { original: filename, folder: '🖼️ /Media_Cleanup', action: 'Queued 30-day auto-archive' };
} else if (lower.includes('unsubscribe') || lower.includes('eml') || lower.includes('free_gift')) {
return { original: filename, folder: '🗑️ /Trash_Bin', action: 'Marked for local purge' };
}
return { original: filename, folder: '📂 /Archive/Unsorted', action: 'Cleaned filename' };
});
}

How I Built It

As a first-year student working solo with a zero-dollar cloud budget, I designed ClutterSloth around a minimalist single-page layout:

Input & Control Panel: A clean dark-mode (#0F172A) textarea pre-loaded with sample desktop and inbox clutter, paired with an interactive Local RAM Mode (Offline) indicator showing an open model loaded directly in memory.
Deterministic State Transition: Clicking Run Local AI Sort Loop transitions the right structured output panel from Awaiting Run to categorized folder groups (/Finance_Receipts, /Work_Notes, /Media_Cleanup, /Trash_Bin).
Responsive Auto Layout: Built with fluid flex constraints that stack cleanly from a 2-column desktop workspace into a mobile-friendly vertical flow.

Why Does Open Innovation Matter?

Client-side privacy: Your filenames and inbox subjects stay strictly inside your browser tab and never touch a remote server, eliminating third-party data exposure.
Zero-cost sustainability: Removing recurring API subscription fees makes practical automation accessible to students, classrooms, and offline environments.
Rule-based efficiency: Small, inspectable logic loops are fast, deterministic, and easy for any developer to audit and extend.

My Agent Session

• Session Workflow: Prototyped and iterated via interactive text-to-code design loops, validating offline state transitions and zero-network privacy constraints.

The Lazy Payoff

By replacing manual renaming with a single-click local loop, ClutterSloth turns a 12-minute weekly desktop cleanup chore into a 0-second background habit—leaving more time to actually sleep or code.

# MenuInsight AI — Decoding Restaurant Menus Locally with Open AI

SUMEDHA GHOSH — Sun, 18 Oct 2026 09:09:30 +0000

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

What I Built

MenuInsight AI is a smart, accessible web tool built for food-loving tourists, diners with severe dietary restrictions, and local food enthusiasts. It lets users type or paste traditional, foreign, or complex dish names to receive instant translations, deep cultural context, detailed allergen tracking flags, and localized guidance on how to authentically enjoy the meal.

As a first-year student living in a residential hostel, I realized how difficult it is to navigate unfamiliar menus when exploring new city food hubs. MenuInsight AI solves this exact problem by acting as a pocket-sized culinary tour guide. It is a tool that travelers, food bloggers, and local independent restaurants would easily pay a subscription for to enhance structural menu accessibility.

Demo

Experience the application live here: Live MenuInsight AI Web Link

Code

The underlying project structural layout, frontend stylesheets, and core prompt workflows are fully open-source.
Check out the repository here: MenuInsight AI GitHub Repository

How I Built It

The prototype leverages lightweight open-weight models optimized for quick local text inferences. By focusing on prompt engineering frameworks, the interface accepts input queries (such as Misal Pav, Tteokbokki, or Pastel de Nata) and routes them into structured semantic outputs.

The entire framework was built utilizing HTML5, CSS Flexbox layouts for mobile responsiveness, and client-side JavaScript execution loops to generate dynamic "AI Insight Cards" right in the viewport without bulky database dependencies.

Why Does Open Innovation Matter?

Open innovation directly protects user privacy while driving down production overhead. Traditional closed-source options require expensive, pay-per-request API queries that force developers to push high costs onto end consumers or small local restaurant partners.

By building MenuInsight AI around open frameworks and edge-based local capabilities:

Zero Runtime Tolls: The open pieces handle inference demands locally, meaning running the core engine costs nothing to scale.
Complete Privacy: Diners can process sensitive personal health criteria, like life-threatening allergen configurations, without transferring their personal health data onto external tracking servers they cannot control.
No Vendor Lock-in: Small businesses avoid getting trapped in unpredictable, corporate-controlled pricing loops, keeping the platform perpetually affordable.

My Agent Session

To review the design process, interface iterations, and architectural logic loops handled during building sessions, you can view the complete blueprint log sequence here: MenuInsight AI Agent Configuration Link

Prize Categories

Featured Categories: Side Hustle / Open-Weight AI Utility

EcoRoute: An Interactive Full-Stack React Application for Environmental Navigation and Green Travel Analytics

SUMEDHA GHOSH — Sat, 10 Oct 2026 17:35:31 +0000

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

EcoRoute is an interactive, full-stack navigation utility designed to reverse standard travel metrics. While conventional routing tools optimize solely for raw speed, EcoRoute actively routes travelers through tree-lined passages, canopy coverage, and localized public parklands.

The application turns outdoor transit into an engaging activity using the "Selfie Wall" and "Proof of Grass Contact" components, allowing travelers to log their route journeys and verify live environmental immersion.

TL;DR

-What: EcoRoute leverages mapping elements to turn standard city transit into nature journeys, prompting commuters to look away from their displays and engage with local green zones.
-The Core AI Strategy: Utilizes highly efficient local inference architectures to process unstructured traveler field logs, converting text or voice notes into clean JSON metadata arrays.

Cost & Footprint: Powered entirely by open-weight, edge-optimized models to enable zero-cost client-side processing without expensive, heavy server configurations.
Everything is Open: Front-end assets, styling components, and local environment structures are public and accessible to all developers.

3 Findings in 30 Seconds

Canopy-first path routing decreases mental exhaustion: Prioritizing side streets with heavy tree lines over high-traffic main asphalt lines offers a significantly more calming journey profile.
Dynamic UI layout keeps interactions brief: Incorporating clean CSS Flexbox design and modular React components makes logging real-time views rapid, keeping terminal use to a minimum.
Local AI configurations match complex hosted APIs: Running small, edge-optimized local processing setups completely eliminates ongoing operational subscription costs while safely protecting sensitive user trip logs.

Demo

Live Production Application: https://ecoroute-touch-grass.emergent.host
Interactive Prototyping Link: https://emergentagent.com

Code

The entire operational repository, layout structure, and components can be reviewed here:

sumedhaghosh302008 / Ecoroute

An eco-friendly routing and navigation prototype built for the Hacktoberfest Open-Source AI Challenge (Week 1: Touch Grass).

Ecoroute

An eco-friendly routing and navigation prototype built for the Hacktoberfest Open-Source AI Challenge (Week 1: Touch Grass).

View on GitHub

How I Built It

As a first-year B.Tech computer science engineering student expanding my development skillset, I designed and deployed this full-stack project using modular modern workflows:

Architecture Core: Built as a scalable React single-page web app utilizing streamlined configuration maps for responsive UI component rendering.
Interface Prototyping: Modeled and fast-tracked using advanced generative design canvases on the Emergent workspace to balance style and mobile accessibility.
Local AI Pipeline: Configured with clean structured parameters to support local open-weight model integration (such as local Gemma 2B instances) to parse destination lists into dependable data structures.

Why Does Open Innovation Matter?

Open-source design is vital for community-driven environmental tools. Proprietary mapping engines block local trail data behind steep platform API call fees and restrictive enterprise licensing. By keeping EcoRoute open-source, local trail groups, municipal bodies, and student developers worldwide can freely download, expand, and personalize the route mapping code to emphasize regional parks and community spaces without overhead.