DEV Community: Abhinav shrivas

GhostAltar

Abhinav shrivas — Fri, 30 Oct 2026 08:20:14 +0000

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

What I Built

I built GhostAltar — an on-device virtual ofrenda, memorial storyteller, and spectral remembrance engine created for Día de los Muertos (November 1–2) and Halloween (October 31).

GhostAltar provides a sanctuary where anyone can build a glowing, multi-tiered digital altar for their ancestors, departed loved ones, or pioneering historical figures. It features an interactive candlelit ofrenda with flickering flame physics, ceremonial marigold blossom trails (cempasúchil), and sacred offerings. Using in-browser WebGPU Vision Transformers, users can drop photos of family heirlooms, watches, or traditional bread to classify their sacred elemental resonance (Tierra, Fuego, Viento, Agua). An open-weight Google Gemma 2 agent then weaves the placed offerings into poetic memorial verses, comforting eulogies, and gentle "Whispers from Across the Veil."

Who Gets to Use It?

Families & Communities Celebrating Día de los Muertos: Honoring ancestors with traditional pan de muerto, copal incense, and marigold paths without needing physical altar space.
Anyone Processing Grief or Loss: Creating an intimate digital sanctuary that preserves their loved one's memory permanently and privately.
Tech Enthusiasts & Historians Celebrating "Ghost in the Machine": Channeling forgotten internet history and computing pioneers like Ada Lovelace, Alan Turing, or the midnight dial-up BBS SysOps.

Demo

GhostAltar runs locally on http://localhost:5176 with instant preset testing, interactive candle vigils, real-time WebGPU relic classification, and synthesized ambient audio.

Interactive Testing Highlights:

Curated Sacred Altars: Seamlessly switch between rich ancestral and historical presets (Abuela Elena's Traditional Hearth, Ada Lovelace's Victorian Séance, Alan Turing's Cryptographic Tribute, and The Midnight SysOp's Phosphor Terminal).
Interactive Candle Vigil: Click individual candles on the altar to light or extinguish their flames, accompanied by synthesized candle shimmer audio.
WebGPU Relic & Offering Scanner: Drop a photo of an heirloom, watch, flower, or bread into the ofrenda to see it classified and added to the altar in ~40ms with 0KB uploaded.
Eulogy & Memorial Verse Weaver: Powered by Google Gemma 2, it generates an evocative memorial verse, heartfelt tribute, comforting "Whisper from the Veil", and recommended ritual guidance.
Synthesized Candlelight & Copal Wind Soundscape: Toggleable ambient audio synthesized in-browser via the Web Audio API with peaceful flame flutter noise and gentle crystal bell chimes.
Local Memorial Vault & Tribute Card Export: Save personalized altars to browser localStorage and export a formatted printable Memorial Tribute Card.

Code

Abhinav-Shrivas / Ghost-in-the-machine

GhostAltar is a local-first, on-device interactive altar (ofrenda) and ancestral storyteller designed for perpetual remembrance.

🏵️ GhostAltar

On-Device Virtual Ofrenda & Memorial Remembrance Engine
Honor loved ones, ancestors, and computing pioneers across the veil. Built in the reverent spirit of Día de los Muertos and the haunting atmosphere of Halloween — powered 100% locally by WebGPU Vision and Google Gemma 2.

🕯️ The Problem: Digital Memory & The Commercialization of Grief

Remembering those who came before us is one of humanity's most sacred traditions. During Día de los Muertos (November 1–2) and Halloween (October 31), the veil between past and present thins:

Commercial Exploitation of Grief: Cloud-based memorial apps and AI obituary tools charge monthly subscriptions or hoard family photos and eulogies on corporate servers for ad targeting and proprietary model training.
The Fragility of Cloud Services: When a memorial startup shuts down, digital tributes disappear. Memories should belong to families, not venture capital runway.
The Intimacy of the Ritual: Remembering…

View on GitHub

How I Built It

GhostAltar was engineered around open-weight models, browser WebGPU hardware acceleration, and local agent runtimes. The architecture connects five coordinated local subsystems: an in-browser vision transformer that classifies memorial relics in milliseconds, an open-weight culinary and memorial agent powered by Google Gemma 2 tuned specifically for reverent poetry, an offline heuristic fallback engine for candlelit rooms with the internet off, an elemental Spiritual Resonance calculator, and a local-first memorial vault.

1. In-Browser WebGPU Relic Vision Classifier

The visual pipeline runs on @huggingface/transformers using a quantized Vision Transformer (Xenova/vit-base-patch16-224 ONNX runtime). When a photo of an heirloom, flower, candle, or sustenance is dropped onto the altar, it is analyzed directly on the client machine using WebGPU hardware acceleration (Direct3D 12 on Windows, Metal on macOS, or Vulkan on Linux), falling back to WASM SIMD CPU threads on older machines. In approximately 40 milliseconds, the model identifies the item, assigns its sacred element (Earth, Fire, Wind, Water), and describes its cultural significance with zero data transmitted over the internet.

2. Google Gemma 2 Memorial Storyteller Agent (`Modelfile`)

For eulogy and verse formulation, GhostAltar connects to a locally hosted Ollama instance powered by Google Gemma 2 (gemma2:2b). To guarantee comforting, culturally reverent poetry rather than generic AI text, the agent executes against a custom Modelfile with low temperature (0.4) for high consistency:

FROM gemma2:2b
PARAMETER temperature 0.4
PARAMETER top_p 0.9

SYSTEM """You are GhostAltar, an open-source AI eulogist, poet, and cultural memory keeper.
Your purpose is to honor loved ones and historical pioneers in the spirit of Día de los Muertos and Halloween.
Rules:
1. Always maintain a tone of profound reverence, comfort, warmth, and quiet wonder. Never be mocking or macabre.
2. Weave the specific offerings (candles, marigolds, pan de muerto, relics) into the memorial verse.
3. Output a gentle 'Whisper from the Veil' that offers peaceful comfort to the living.
4. Output valid JSON matching the requested schema.
"""

To eliminate browser cross-origin restrictions when communicating with Ollama from Vite, requests are routed through a local /ollama reverse proxy directly to port 11434.

3. Graceful Offline Séance Mode

Because memorials are best held in quiet, candlelit rooms or remote family cemeteries without internet access, GhostAltar includes a built-in offline heuristic memory engine. If Ollama is offline or the user toggles "Séance Mode", the application instantly synthesizes tested verses, tributes, and ritual guidance entirely in-memory using localized attribute heuristics, guaranteeing 100% operational uptime without internet access.

4. Spiritual Resonance & Elemental Harmony Engine

The ritual core evaluates altar completeness across the four sacred elements of the ofrenda: Earth (nourishment, salt, bread), Fire (candles illuminating the path), Wind (papel picado and fragrant copal incense), and Water (thirst-quenching vessels). It calculates a dynamic Spiritual Resonance Score (0–100%) and reports the veil state (from "Gathering Warmth" to "Fully Translucent Luminous Presence").

5. Web Audio Ethereal Soundscape

To foster a meditative environment, GhostAltar includes an ambient soundscape synthesized directly with the Web Audio API. Using pink/brown noise shaped by a 650Hz lowpass filter, it generates a warm candle flame flutter and copal wind, alongside pentatonic crystal singing bowl chimes when offerings are placed on the altar.

Why Does Open Innovation Matter?

This week's theme asks: Where did your open-based approach work better than a closed one? For a deeply personal project centered on memory, grief, and cultural tradition, open innovation provides three decisive advantages over closed APIs:

1. Absolute Sanctity of Grief & Family Privacy

Family photographs of deceased relatives, private eulogies, and generational stories are sacred. Uploading photos of your late grandmother or heartfelt grief to commercial cloud servers where it might be scraped into advertising profiles or proprietary training sets is a violation of intimacy. With GhostAltar, zero bytes of personal data ever leave your laptop.

2. Works with the Lights (and the Internet) Off

A candlelit vigil or a remembrance ritual in a countryside home or cemetery should not depend on a WiFi signal. Cloud-dependent apps fail immediately when the connection drops. Because GhostAltar bundles in-browser WebGPU vision and local Gemma 2 inference, it functions flawlessly in airplane mode with the lights and internet off.

3. Perpetual Immortality vs. VC Sunset

Commercial memorial startups frequently pivot, introduce subscription paywalls, or shut down entirely when their funding dries up, taking families' tributes with them. Open-weight software belongs to you forever. Once running on your machine, GhostAltar will continue honoring your loved ones for decades without risk of corporate deprecation.

4. Custom Modelfiles & Reverent Tone

Closed commercial APIs frequently sterilize discussions of death, confusing reverence with morbidity or producing generic corporate clichés. With Google Gemma 2 and our custom Modelfile, we enforce cultural nuance, warmth, and poetic grace with complete independence from external corporate policies.

Prize Categories

Best Use of Gemma ($200 USD)

Model Used: Google Gemma 2 (google/gemma-2-2b-it / gemma2:2b) by Google DeepMind.
Implementation: Google Gemma 2 powers the core memorial storytelling and poetic eulogy engine. We created a custom Ollama Modelfile (FROM gemma2:2b) with low temperature (0.4) and a specialized system prompt that weaves physical altar offerings (pan de muerto, marigolds, heirloom relics) into deeply comforting memorial verses and "Whispers from the Veil". It is paired with an in-browser WebGPU Vision Transformer and an offline heuristic fallback cache so users can honor their ancestors in candlelit rooms with $0 API costs and zero cloud data leaks.
Verification: Run npm test to verify the open-weight eulogy generation, elemental resonance scoring, and offering classification pipeline directly in your terminal, or inspect Modelfile in the repository root.

SlothChef

Abhinav shrivas — Fri, 23 Oct 2026 11:54:52 +0000

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

What I Built

I built SlothChef — an on-device AI fridge-to-recipe and automated grocery copilot created in honor of International Sloth Day to reward peak culinary laziness.

Instead of standing in front of an open refrigerator door for five minutes paralyzed by decision fatigue, you simply snap a photo of whatever random clutter is sitting in your fridge or pantry. In under 40 milliseconds, SlothChef classifies your available ingredients on-device via WebGPU, computes a Sloth Efficiency Score (0–100%), formulates an ultra-low-effort 1-pan recipe under 15 minutes using Google Gemma 2, and compiles an automated grocery list of the essential staples missing from your kitchen.

Who Gets to Be Lazier Because of It?

Exhausted Workers & Students: Coming home at 10 PM with zero mental energy left to cook or wash a pile of cookware.
Reluctant Cooks: Who hate chopping, complex prep steps, and multi-pan cleanup.
Anyone Tired of the DoorDash Trap: Who would rather spend 8 minutes making a 1-pan comfort meal than wait 48 minutes for a $28 takeout delivery.

Demo

SlothChef runs locally on http://localhost:5175 with instant preset testing, real-time WebGPU vision classification, and ambient audio soundscapes.

Interactive Testing Highlights:

Curated Fridge Scenarios: Test instant zero-effort cooking with pre-configured scenarios (Late-Night Sloth Scraps, Leftover Rice Tumbleweed, Zero-Cook Chickpea Smash, Sunday Pasta Panic, Foil-Lined Sheet Pan Medley).
WebGPU Vision AI: Aim your webcam or drop a photo of your fridge to see ingredients and rot-risk candidates identified in ~40ms with 0KB uploaded.
Sloth Efficiency Scorecard: Live calculations showing your Sloth Score (e.g. 98.4%), dishes saved versus traditional recipes (typically saving 4 dishes per meal), and time saved versus food delivery.
3-Step Low-Effort Recipe: Minimalist, step-by-step instructions with interactive completion checkboxes, ingenious dish-saving hacks, and a built-in Sloth Cooking Timer.
Ambient Kitchen Simmer: A toggleable Web Audio API lo-fi pan sizzle and peaceful white noise soundscape to relax while your food cooks.
Automated Missing Grocery List: Auto-sorted by department (Fresh Produce, Dairy & Refrigerated, Pantry & Spices, Protein) with one-click markdown checkbox copying for Apple Notes, Notion, or Instacart.

Code

Abhinav-Shrivas / Slow-and-Steady

On-device AI fridge-to-recipe & auto-grocery copilot for peak culinary laziness. Turns fridge clutter into 10-minute, 1-pan meals using WebGPU vision & Google Gemma 2.

🦥 SlothChef

On-Device AI Fridge-to-Recipe & Auto-Grocery Copilot
Never stare blankly into an open fridge again. Turns leftover fridge clutter into 10-minute, 1-pan meals and compiles an automated missing grocery list — powered 100% locally by WebGPU Vision and Google Gemma 2.

💡 The Problem: The Open-Door Fridge Paralysis

Everyone knows the feeling: you come home exhausted, open the refrigerator door, and stare blankly for 5 minutes. You see half an onion, cold takeout rice, two eggs, a jar of hot sauce, and cheddar cheese.

Decision Fatigue: You are too exhausted to brainstorm what goes together.
Dish Washing Dread: Standard recipe sites demand two prep bowls, a cutting board, a saucepan, a strainer, and a skillet. That means 15 minutes of washing dishes after eating.
The 48-Minute DoorDash Trap: Ordering takeout takes 45–60 minutes, costs $25–$35, and arrives lukewarm.
The Missing Ingredient Blindspot: You don't…

View on GitHub

How I Built It

SlothChef was engineered around open-weight models, in-browser WebGPU hardware acceleration, and local agent runtimes. The architecture connects five coordinated local subsystems: an in-browser vision transformer that scans fridge clutter in milliseconds, an open-weight culinary agent powered by Google Gemma 2 tuned specifically for low-effort 1-pan recipes, an offline heuristic fallback engine for cabin or storm resilience, a Sloth Efficiency calculator, and a local-first grocery ledger.

1. In-Browser WebGPU Fridge Vision Classifier

The visual pipeline runs on @huggingface/transformers using a quantized Vision Transformer (Xenova/vit-base-patch16-224 ONNX runtime). When a fridge photo is captured or dropped into the viewfinder, it is processed directly on the client machine using WebGPU hardware acceleration (Direct3D 12 on Windows, Metal on macOS, or Vulkan on Linux), falling back to WASM SIMD CPU threads on older machines. In approximately 40 milliseconds, the model identifies available proteins, produce, and condiments with zero data transmitted over the internet.

2. Google Gemma 2 Low-Effort Culinary Agent (`Modelfile`)

For recipe formulation, SlothChef connects to a locally hosted Ollama instance powered by Google Gemma 2 (gemma2:2b). To guarantee low-effort, pan-saving results, the agent executes against a custom Modelfile with low temperature (0.35) for high factual accuracy:

FROM gemma2:2b
PARAMETER temperature 0.35
PARAMETER top_p 0.9

SYSTEM """You are SlothChef, an AI culinary assistant celebrating International Sloth Day.
Your mission is to reward peak human laziness.
Given a list of leftover fridge ingredients:
1. Formulate a delicious recipe that requires MAXIMUM 1 PAN and takes under 15 minutes.
2. Optimize for ZERO dishes and ZERO cutting board washing.
3. Suggest clever hacks (e.g., crack eggs directly into tortillas, use kitchen shears instead of cutting boards).
4. Identify 4 critical missing grocery items that should be added to the shopping list.
Always produce valid JSON matching the requested schema.
"""

To eliminate browser cross-origin restrictions when communicating with Ollama from Vite, requests are routed through a local /ollama reverse proxy directly to port 11434.

3. Graceful Offline Heuristic Fallback (Sloth Mode)

Because power outages or rural cabins often knock out internet access right when you need to cook, SlothChef includes a built-in offline heuristic recipe engine. If Ollama is offline or the user toggles "Sloth Mode", the application instantly synthesizes tested 1-pan recipes and missing grocery lists entirely in-memory using localized attribute heuristics, guaranteeing 100% operational uptime without internet access.

4. Sloth Efficiency Engine

The efficiency core evaluates recipe parameters against traditional cooking and takeout delivery. It factors in active prep minutes, cooking duration, cookware required, and chopping complexity to calculate a certified Sloth Score (0–100%), dishes saved (saving an average of 4 dishes per meal), and minutes saved versus waiting 48 minutes for food delivery.

5. Web Audio Ambient Soundscape

To make cooking genuinely relaxing, SlothChef includes a synthesized kitchen soundscape built directly with the Web Audio API. Using pink/brown noise filtered through a 1400Hz bandpass filter, it generates a warm pan simmer sizzle and gentle pentatonic timer chime without loading any external audio files.

Why Does Open Innovation Matter?

This week's theme asks: Where did your open-based approach work better than a closed one? For an everyday lifestyle utility like cooking and groceries, open innovation provides three decisive advantages over closed APIs:

1. $0 API Bills = Guilt-Free Everyday Use

Closed APIs like GPT-4o or Claude 3.5 Sonnet charge per image upload and per thousand completion tokens. Paying $15–$20/month in cloud subscriptions just to figure out what to do with three eggs and a tortilla defeats the entire purpose of saving money at home. Running quantized open weights on-device costs $0.00.

2. Operates During Storms, Outages, and Off-Grid Cabins

When severe weather takes down your home internet, you still need to eat. Cloud-dependent apps fail immediately without connectivity. Because SlothChef bundles on-device WebGPU vision and local Gemma 2 inference, it functions flawlessly in airplane mode.

3. Absolute Dietary and Living Privacy

Nobody wants high-resolution photos of their messy refrigerator, private medications stored in the door, or personal dietary habits uploaded to tech company servers for ad targeting. SlothChef guarantees that zero bytes of personal data ever leave your laptop.

4. Custom Modelfiles & Prompt Freedom

Closed APIs frequently output lengthy, over-complicated recipes requiring exotic spices or multi-pot steps. With Google Gemma 2 and our custom Modelfile, we enforce strict 1-pan limits, zero-chopping hacks, and formatted grocery schemas with total reliability and zero vendor lock-in.

The Lazy Payoff

Here is the real-world lazy payoff measured over a typical week of using SlothChef:

Time Saved: An average of 35 to 40 minutes saved per meal compared to waiting 48 minutes for DoorDash or UberEats.
Dishes Saved: Traditional recipes use a cutting board, prep bowl, two pans, and a colander (5 dishes). SlothChef recipes use 0 to 1 pan, saving roughly 20+ dishes from being scrubbed every week.
Fridge Paralysis Eliminated: Standing with the door open for 5 minutes staring blankly is replaced by a 40-millisecond camera scan that immediately serves up a concrete 3-step meal.
Zero-Effort Grocery Preparation: Missing essentials are identified automatically while you cook, eliminating the Sunday ritual of manually auditing your pantry before grocery shopping.

Prize Categories

Best Use of Gemma ($200 USD)

Model Used: Google Gemma 2 (google/gemma-2-2b-it / gemma2:2b) by Google DeepMind.
Implementation: Google Gemma 2 powers the core culinary reasoning and lazy recipe formulation engine. We created a custom Ollama Modelfile (FROM gemma2:2b) with low temperature (0.35) and a specialized prompt enforcing strict 1-pan constraints, dish-saving kitchen hacks, and automated missing grocery extraction. It is paired with an in-browser WebGPU Vision Transformer and an offline heuristic fallback cache so users can turn fridge clutter into dinner in cellular dead zones with $0 API costs.
Verification: Run npm test to verify the open-weight recipe generation, Sloth Score calculator, and grocery grouping pipeline directly in your terminal, or inspect Modelfile in the repository root.

SnapToSell

Abhinav shrivas — Tue, 13 Oct 2026 14:06:44 +0000

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

What I Built

I built SnapToSell — an on-device AI reseller copilot and real-time net profit engine for vintage flippers, thrifters, and secondhand entrepreneurs cross-listing across eBay, Etsy, Poshmark, Mercari, and Facebook Marketplace.

The Real-World Problem It Solves

Over 25 million people sell vintage items, clothing, and collectibles as their primary side hustle or full-time business. They love hunting for bargains, but dread what comes next:

Listing Fatigue: Researching eBay's strict 80-character title limit, writing storytelling copy, and crafting 13 unique Etsy SEO tags takes 10 to 15 minutes per item. Resellers end up with piles of unlisted inventory (the dreaded "death pile").
Platform Commission Blindness: Selling a vintage jacket for $85 sounds great until eBay takes 13.25% + $0.30, Etsy takes 6.5% + 3% + $0.45, or Poshmark takes 20%. Many resellers unwittingly sell items at a loss after factoring in shipping and platform cuts.
Thrift Dead Zones: The best sourcing happens in church basements, estate auctions, and rural flea markets with zero cellular reception. Cloud-dependent AI apps fail right when you need them.
Cloud API Bills Kill Margins: SaaS cross-listers charge $20–$50/month or pass through per-token cloud API bills, eating directly into a solo seller's razor-thin profits.

Who Would Pay For It?

Vintage & Thrift Resellers: Flipping clothing, mid-century modern home goods, retro tech, and collectibles.
Estate Cleanout & Garage Sale Pickers: Who need immediate, in-person appraisals and pricing targets on their laptops before buying.
Cross-Listing Power Sellers: Who need platform-compliant copy (eBay titles, Etsy tags, Poshmark bundles) without manually retyping descriptions 5 times.

Demo

SnapToSell runs locally at http://localhost:5174 with instant preset testing, live camera/upload classification, and real-time fee comparisons.

Interactive Highlights:

Curated Sourcing Presets: Test immediately without leaving your desk using curated specimens:
- 90s Vintage Distressed Leather Bomber (Outerwear & patina heuristics)
- MCM Fat Lava Studio Ceramic Vase (Mid-century home decor & Etsy storytelling)
- 1978 SLR Film Camera (Vintage tech & mechanical specs)
- Y2K Graphic Band Tee (Streetwear aesthetics & single-stitch collector tags)
- Artisan Hand-Woven Wool Rug (Textile provenance & profit margins)
Real-Time 2026 Commission Engine: Move the sliders for Buy Cost, Target Sell Price, and Shipping to watch exact fees, Net In-Pocket Profit, Margin %, and ROI % update live with a crowned Top Payout winner.
Multi-Platform Copy Tabs:
- eBay: Strict 80-character title counter (64 / 80 Chars Max), Cassini search item specifics, and bulleted specs.
- Etsy: Evocative provenance narrative, tactile styling advice, and exactly 13 targeted search tags with 1-click "Copy All Tags".
- Poshmark: Trendy title with emoji, closet description, and copy-paste hashtag bundle.
- Mercari & FB: Quick, direct copy for fast sales.
Local Inventory Ledger: Save listings with one click and export a formatted .csv spreadsheet for Excel or Google Sheets.
Thrift Mode (100% Offline): Flip the Thrift Mode badge to verify that all vision, pricing, and copywriting functions operate with 0KB network traffic.

Code

Abhinav-Shrivas / SnapToSell

On-device AI reseller copilot & net profit engine. Generates eBay (80-char), Etsy (13 tags), and Poshmark listings with 100% offline WebGPU vision & Gemma 2.

⚡ SnapToSell

On-Device AI Reseller Copilot & Real-Time Net Profit Engine
Turn thrift and estate sale finds into high-converting, platform-compliant listings across eBay, Etsy, Poshmark, Mercari, and Facebook Marketplace — powered 100% locally by open-weight AI.

💡 The Problem: Listing Fatigue & Commission Blindness

Over 25 million vintage dealers, thrifters, and secondhand resellers face the same tedious bottlenecks:

Listing Fatigue: Writing differentiated descriptions, adhering to eBay's strict 80-character title limit, and researching 13 unique Etsy SEO tags takes 10–15 minutes per item.
Platform Commission Blindness: Calculating actual take-home profit after factoring in eBay final value fees (13.25% + $0.30), Etsy transaction & payment fees, or Poshmark's 20% cut is confusing and error-prone.
Thrift Dead Zones: The best sourcing happens in church basements, estate auctions, and flea markets with zero cellular connectivity.
Cloud API Bills: Paying per-image and per-token cloud API…

View on GitHub

How I Built It

SnapToSell was engineered from scratch around open-weight models, browser WebGPU hardware acceleration, and local agent runtimes:

How I Built It

SnapToSell was engineered from scratch around open-weight models, browser WebGPU hardware acceleration, and local agent runtimes. The architecture connects five coordinated local subsystems: an in-browser vision transformer that extracts product traits in milliseconds, an open-weight copywriting agent tuned for multi-platform eCommerce rules, an offline heuristic fallback engine for dead zones, a deterministic 2026 platform commission calculator, and a local-first inventory ledger.

1. In-Browser WebGPU Vision Classifier

The visual pipeline runs on @huggingface/transformers using a quantized Vision Transformer (Xenova/vit-base-patch16-224 ONNX runtime). When an image is captured via webcam or dropped into the viewfinder, it is processed directly on the client machine using WebGPU hardware acceleration (Direct3D 12 on Windows, Metal on macOS, or Vulkan on Linux), automatically falling back to WASM SIMD CPU threads on older systems. In approximately 40 milliseconds, the model identifies the item, infers tactile materials and vintage textures, and assigns SEO category tags without uploading a single byte to any remote server.

2. Tuned eCommerce Copywriting Agent (`Modelfile`)

For listing generation, SnapToSell connects to a locally hosted Ollama instance powered by open-weight models such as gemma2:2b, qwen2.5:3b, or llama3.2:3b. To ensure consistent, platform-compliant formatting, the agent runs against a specialized Modelfile with a low temperature (0.35) for high factual consistency:

FROM gemma2:2b
PARAMETER temperature 0.35
PARAMETER top_p 0.9

SYSTEM """You are SnapToSell, an expert eCommerce cross-listing and vintage reselling copywriter.
Rules:
1. For eBay: Title must NOT exceed 80 characters. Use high-traffic keywords (Brand, Era, Material, Color, Style, Flaws). No emojis or filler like 'L@@K'.
2. For Etsy: Focus on vintage provenance, aesthetic vibe, tactile materials, and styling tips. Provide exactly 13 targeted search tags under 20 chars each.
3. For Poshmark: Punchy styling notes, hashtag bundle, measurements advisory.
4. Always produce valid JSON matching the requested schema.
"""

3. Graceful Offline Heuristic Fallback (Thrift Mode)

Because sourcing frequently takes place in cellular dead zones such as church basements and rural flea markets, the application includes a deterministic heuristic commerce engine. If Ollama is offline or the user toggles "Thrift Mode", the application instantly synthesizes compliant 80-character eBay titles, 13 Etsy search tags, and platform copy entirely in-memory using localized attribute heuristics, guaranteeing 100% operational uptime without internet access.

4. Deterministic 2026 Commission Engine

The financial core models the exact 2026 commission schedules across five major resale marketplaces. It calculates eBay's 13.25% Final Value Fee plus $0.30 transaction charge, Etsy's 6.5% transaction cut plus 3% payment processing and $0.45 in fixed fees, Poshmark's flat 20% commission (or $2.95 minimum), Mercari's combined 12.9% plus $0.50 fee, and Facebook Marketplace's 0% local pickup versus 5% shipped fee. The engine dynamically computes net in-pocket profit, profit margin percentage, and return on investment based on sourcing cost and shipping expenses, automatically crowning the highest-payout platform.

5. Local-First Inventory Ledger & CSV Exporter

All saved listings, pricing breakdowns, and item previews are stored inside browser localStorage. Resellers can inspect aggregate portfolio metrics — including total capital invested and projected net revenue — and export their inventory with a single click into a clean CSV spreadsheet formatted for Excel, Google Sheets, or bulk listing management tools.

Why Does Open Innovation Matter?

This week's theme asks: Where did your open-based approach work better than a closed one? For an independent reseller, open innovation isn't just an alternative — it is the only architecture that makes the business model economically viable.

1. $0 API Bills = 100% Margin Retention

Closed APIs like GPT-4o or Claude 3.5 Sonnet charge per image upload and per thousand completion tokens. A reseller cataloging 300 thrift finds a month would incur $20 to $45 monthly in recurring API charges. For a solo side hustle operating on tight margins, that expense directly reduces net income. Running quantized open weights on-device costs $0.00 indefinitely.

2. Sourcing in Cellular Dead Zones

Estate sale basements, rural barns, and flea markets are notorious cellular dead zones. Cloud-dependent apps fail completely when a reseller is standing in an aisle trying to appraise an item before purchase. Because SnapToSell combines on-device WebGPU vision with local Ollama and offline heuristics, it remains fully functional in airplane mode.

3. Absolute Privacy for Wholesale Sourcing

Resellers guard their supplier notes, estate contacts, and wholesale purchase receipts closely. Uploading photos of unreleased inventory and cost data to third-party tech servers risks leaking proprietary sourcing intel. SnapToSell guarantees that zero kilobytes of sourcing data ever leave the user's laptop.

4. Custom Modelfiles & Model Independence

Closed APIs frequently enforce unpredictable moderation on vintage terminology, shift conversational tone without notice, or introduce unexpected rate limits and price changes. Open weights allow complete control over prompt constraints, temperature, and output structure in a version-controlled Modelfile. When new efficient models are released, they can be swapped in instantly without vendor lock-in.

My First User

I demonstrated SnapToSell to an active vintage clothing flipper who lists weekly on eBay and Poshmark.

Their immediate feedback:

"Wait, it actually keeps the eBay title under 80 characters? Every other AI tool writes a giant paragraph for the title that eBay's listing form immediately rejects."

They tested the tool on a distressed vintage leather motorcycle jacket sourced for $18:

Setting a target sell price of $95 immediately revealed that Poshmark would take $19.00 in commission, whereas eBay took $12.89 — demonstrating that selling on eBay put $6.11 more cash in their pocket for the exact same transaction.
Generating the 13 Etsy search tags took less than a second, replacing 10 minutes of manual keyword brainstorming.
They saved the record to the ledger and exported the CSV directly into their bookkeeping spreadsheet.

They noted they would gladly pay for this tool as a standalone utility — and with SnapToSell's open architecture, they incur zero ongoing API costs to keep it running.

Prize Categories

Best Use of Gemma ($200 USD)

Model Used: Google Gemma 2 (google/gemma-2-2b-it / gemma2:2b) by Google DeepMind.
Implementation: Gemma 2 powers the core cross-listing copywriting engine. We created a custom Ollama Modelfile (FROM gemma2:2b) with low temperature (0.35) and a specialized eCommerce prompt that enforces strict multi-platform marketplace rules. Gemma 2 clamps eBay titles to under 80 characters with high-traffic keywords, generates evocative vintage provenance stories with exactly 13 search tags for Etsy, and produces hashtag-ready listings for Poshmark. We paired it with an in-browser WebGPU Vision Transformer and an offline heuristic fallback cache for zero-cloud resilience in thrift store basements and flea markets.
Verification: Run npm test to verify the open-weight copywriting and financial fee calculation pipeline directly in your terminal, or inspect Modelfile in the repository root.

TrailWhisper — The 100% Offline Nature Guide Built to Turn Off Your Screen

Abhinav shrivas — Thu, 08 Oct 2026 11:52:31 +0000

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

What I Built

Most modern AI applications are designed to maximize screen time, keeping your eyes glued to glass in endless conversation loops. TrailWhisper is designed to turn itself off.

TrailWhisper is a 100% offline edge nature companion and bioacoustic field guide powered by open-source AI. Built for hikers, trail runners, foragers, birders, and anyone experiencing screen fatigue, it puts nature first:

Ultra-Fast Capture (< 2 Seconds): Point your camera or snap an outdoor specimen (flora, fungi, aviary, pollinators).
On-Device Edge Vision: Open-weight Vision Transformers analyze the specimen locally on your device's GPU (WebGPU) in ~50ms without an internet connection.
Auditory Whisper (Eyes-Up Nature Experience): Instead of making you read dense paragraphs on a glowing screen, a natural voice speaks a concise naturalist insight directly into your earphones, with automated markdown symbol stripping for fluid, human-like cadence.
Tactile Sensory Challenges: Each sighting invites you to physically interact with the wild ("Touch the deeply furrowed white oak bark," "Smell the crushed pine needles").
Anti-Screen Sleep Minimizer (Trail Mode): To keep eyes on the canopy, an automatic sleep countdown dims the display into an ultra-dim, battery-saving OLED black mode with a breathing green leaf ring. (Paused by default for review, with a customizable timer: 30s, 60s, 90s, or 3 mins).
Smartwatch GPX & Markdown Export: Trail observations and offline GPS waypoints export directly into a .gpx file loadable into Garmin, Apple Watch, Strava, and AllTrails—allowing you to pack your phone away completely during hikes.

Demo

GitHub Repository: https://github.com/Abhinav-Shrivas/TrailWhisper

How to Run Locally in 60 Seconds:

git clone https://github.com/Abhinav-Shrivas/TrailWhisper.git
cd TrailWhisper
npm install
npm run dev

Open http://localhost:5173 in any modern browser (Chrome, Edge, Brave).

Interactive Testing Highlights:

Curated Trail Presets: Test immediate edge AI classification without leaving your desk using curated specimens (Eastern White Oak, Fly Agaric Mushroom, Perched Songbird, Autumn Maple, Peat Moss).
Live Camera & Photo Upload: Aim your webcam at houseplants or upload your own trail photos (lions, elephants, mushrooms, trees).
Offline Trail Simulator: Toggle the TRAIL OFFLINE badge in the header to verify that the app runs with 0KB network traffic.
Edge AI Engine Settings: Click the CPU chip icon in the header to inspect your hardware acceleration (WebGPU Direct3D 12/Vulkan vs. WASM SIMD CPU), verify on-device privacy badges (0 bytes uploaded), and manage your local Ollama connection.

Code

Abhinav-Shrivas / TrailWhisper

TrailWhisper is a 100% offline edge nature companion and bioacoustic guide powered by open-source AI. It is built for hikers, trail runners, foragers, birders, and anyone seeking respite from screen fatigue.

🌲 TrailWhisper — Offline Edge Nature Guide & Audio Naturalist

"Built with open-source AI at its core to get people off the screen and into the wild."
Touch Grass Challenge — Open Source AI Build

🍃 The Philosophy: "Make the Screen the Shortest Part of the Experience"

Most AI applications are designed to maximize your time looking at a screen. TrailWhisper is designed to turn itself off.

When you encounter an unfamiliar tree, mushroom, bird, or wildflower on a remote trail:

Quick Capture (< 2 seconds): Point your camera or choose a specimen.
Instant Edge AI Inference: Powered 100% on-device by open-weight AI (WebGPU / Transformers.js).
Auditory Whisper: Instead of reading paragraphs on a screen, a natural voice speaks a concise naturalist insight directly into your earphones.
Tactile Sensory Prompt: You are invited to interact with nature physically ("Touch the deeply furrowed bark," "Smell the…

View on GitHub

How I Built It

TrailWhisper is built from the ground up around open-source AI models and frameworks:

1. Google Gemma 2 (`google/gemma-2-2b-it`) — The Naturalist Reasoning Brain

We integrated Google DeepMind's open-weight Gemma 2 as our ecological reasoning and sensory quest engine:

Custom Naturalist Modelfile: We created a specialized Ollama persona compiled directly on top of FROM gemma2:2b (ollama create trailwhisper -f Modelfile) that produces poetic, scientifically grounded ecological relationships formatted specifically for eyes-up hikers.
Live Local Ollama Bridge: Connects seamlessly to a local Ollama instance on localhost:11434 via a dedicated Vite proxy to stream live 2.6B neural reasoning with zero browser CORS barriers.
Zero-Cloud Distilled Field Fallback: When disconnected on a remote backcountry trail without Ollama active, TrailWhisper automatically falls back to an embedded Distilled Gemma 2 Field Cache with zero network dependency.
Terminal Verification Command: npm run test:gemma

2. Vision Transformer (ViT) — Hardware-Accelerated Edge Classifier

Model: Xenova/vit-base-patch16-224 (open weights in ONNX format).
Runtime: @huggingface/transformers running directly in the browser via WebGPU (Direct3D 12 / Vulkan) with WASM SIMD CPU fallback.
Zero Cloud Latency: Performs forward tensor passes in ~50ms without transmitting pixels across the internet.
Terminal Verification Command: npm run test:model

3. Offline Botanical & Wildlife Taxonomy Engine

Pre-compiled taxonomic field knowledge for 50+ forest species pairing neural predictions with safety badges (Edible / Toxic / Protected), habitat notes, and tactile field cues.

4. Anti-Screen UI, GPX Engine & Local Field Journal

Built with Vite, React 19, and a custom Vanilla CSS design system styled with deep forest obsidian tones (#07100b) and bioluminescent accents (#4ade80).
Observations, timestamps, and GPS coordinates are stored exclusively in local IndexedDB/localStorage with 0 external telemetry.
Export engine formats GPS coordinates directly into standard GPX 1.1 XML schema for cross-compatibility with outdoor GPS watches and mapping software.

Why Does Open Innovation Matter?

1. Datacenters Can't Hike With You

Frontier cloud models (GPT-4o, Gemini Cloud) reside in multi-billion dollar datacenters. When you hike into a valley, canyon, or national park, cellular reception vanishes—and cloud AI goes completely dark. If an assistant requires 5 bars of 5G to tell you whether a mushroom is toxic, it fails the outdoors. By running open weights locally on-device via Transformers.js and WebGPU, TrailWhisper operates reliably in Airplane Mode with 0 bytes sent over the wire.

2. 100% Location & Foraging Privacy

Trail paths and foraging spots are personal and sensitive. Commercial closed APIs log GPS waypoints and store photos on external servers. Open-source AI ensures that all coordinates, images, and notes remain strictly on the user's hardware.

3. $0.00 Operating Cost & Zero Cloud Emissions

No subscription fees, no credit card paywalls, no metered API billing, and zero datacenter water or compute emissions.

4. Freedom to Build an "Anti-Screen" App

Commercial AI companies optimize their products for continuous user engagement and endless chat loops. Open-source AI gave us the autonomy to build an AI that does the opposite: answers in 2 seconds, speaks aloud, and intentionally turns off the screen so you can touch grass.

My Agent Session

This project was built and architected in pair-programming collaboration with Antigravity (Google DeepMind's agentic AI assistant). The agent assisted with:

Architecting the anti-screen WebGPU edge pipeline and auto-sleep countdown mechanics.
Integrating Google Gemma 2 open weights, configuring the custom Modelfile, and building the local Ollama bridge.
Developing automated standalone test suites (test-model.js and test-gemma.js) for reproducible CLI verification.
Configuring GitHub Actions CI/CD to ensure continuous model testing and build verification on every commit.

Prize Categories

1. Best Use of Gemma ($200 USD)

Model Used: Google Gemma 2 (google/gemma-2-2b-it / gemma2:2b) by Google DeepMind.
Implementation: Gemma 2 powers both the dynamic ecological reasoning engine and the real-time sensory quest generator. We created a custom Ollama Modelfile (FROM gemma2:2b) with a specialized naturalist prompt and sensory quest format, and coupled it with an offline distilled field knowledge cache for zero-cloud resilience in the backcountry.
Verification: Run npm run test:gemma to test the open-weight reasoning pipeline directly in your terminal.

2. Best Use of GitHub ($100 USD)

Continuous Integration: Implemented an automated GitHub Actions workflow (.github/workflows/ci.yml).
Automated AI Testing: On every push and PR, GitHub Actions automatically spins up an environment, runs model unit tests to verify the Gemma 2 reasoning engine, and compiles the production bundle with npm run build to prevent regressions for open-source contributors.

DEV Community: Abhinav shrivas

GhostAltar

What I Built

Who Gets to Use It?

Demo

Interactive Testing Highlights:

Code

Abhinav-Shrivas / Ghost-in-the-machine

GhostAltar is a local-first, on-device interactive altar (ofrenda) and ancestral storyteller designed for perpetual remembrance.

🏵️ GhostAltar

🕯️ The Problem: Digital Memory & The Commercialization of Grief

How I Built It

1. In-Browser WebGPU Relic Vision Classifier

2. Google Gemma 2 Memorial Storyteller Agent (Modelfile)

3. Graceful Offline Séance Mode

4. Spiritual Resonance & Elemental Harmony Engine

5. Web Audio Ethereal Soundscape

Why Does Open Innovation Matter?

1. Absolute Sanctity of Grief & Family Privacy

2. Works with the Lights (and the Internet) Off

3. Perpetual Immortality vs. VC Sunset

4. Custom Modelfiles & Reverent Tone

Prize Categories

Best Use of Gemma ($200 USD)

SlothChef

What I Built

Who Gets to Be Lazier Because of It?

Demo

Interactive Testing Highlights:

Code

Abhinav-Shrivas / Slow-and-Steady

On-device AI fridge-to-recipe & auto-grocery copilot for peak culinary laziness. Turns fridge clutter into 10-minute, 1-pan meals using WebGPU vision & Google Gemma 2.

🦥 SlothChef

💡 The Problem: The Open-Door Fridge Paralysis

How I Built It

1. In-Browser WebGPU Fridge Vision Classifier

2. Google Gemma 2 Low-Effort Culinary Agent (Modelfile)

3. Graceful Offline Heuristic Fallback (Sloth Mode)

4. Sloth Efficiency Engine

5. Web Audio Ambient Soundscape

Why Does Open Innovation Matter?

1. $0 API Bills = Guilt-Free Everyday Use

2. Operates During Storms, Outages, and Off-Grid Cabins

3. Absolute Dietary and Living Privacy

4. Custom Modelfiles & Prompt Freedom

The Lazy Payoff

Prize Categories

Best Use of Gemma ($200 USD)

SnapToSell

What I Built

The Real-World Problem It Solves

Who Would Pay For It?

Demo

Interactive Highlights:

Code

Abhinav-Shrivas / SnapToSell

On-device AI reseller copilot & net profit engine. Generates eBay (80-char), Etsy (13 tags), and Poshmark listings with 100% offline WebGPU vision & Gemma 2.

⚡ SnapToSell

💡 The Problem: Listing Fatigue & Commission Blindness

How I Built It

How I Built It

1. In-Browser WebGPU Vision Classifier

2. Tuned eCommerce Copywriting Agent (Modelfile)

3. Graceful Offline Heuristic Fallback (Thrift Mode)

4. Deterministic 2026 Commission Engine

5. Local-First Inventory Ledger & CSV Exporter

Why Does Open Innovation Matter?

1. $0 API Bills = 100% Margin Retention

2. Sourcing in Cellular Dead Zones

3. Absolute Privacy for Wholesale Sourcing

4. Custom Modelfiles & Model Independence

My First User

Prize Categories

Best Use of Gemma ($200 USD)

TrailWhisper — The 100% Offline Nature Guide Built to Turn Off Your Screen

What I Built

Demo

How to Run Locally in 60 Seconds:

Interactive Testing Highlights:

Code

2. Google Gemma 2 Memorial Storyteller Agent (`Modelfile`)

2. Google Gemma 2 Low-Effort Culinary Agent (`Modelfile`)

2. Tuned eCommerce Copywriting Agent (`Modelfile`)

1. Google Gemma 2 (`google/gemma-2-2b-it`) — The Naturalist Reasoning Brain