DEV Community: Waheedur Rahman

GHOST IN THE MACHINE

Waheedur Rahman — Fri, 30 Oct 2026 13:04:57 +0000

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

What I Built

Ghost in the Machine (NecroOS & The Digital Ofrenda) is an open-source, client-side cyber-gothic workstation and sacred memorial altar. It bridges the thin boundary between Halloween (spectral entities, possessed terminals, and eerie nocturnal anomalies) and Día de los Muertos (cempasúchil marigolds, flickering candlelight, and celebrating the memories of loved ones and pioneers who shaped our world).

Core Interactive Modules

The Haunted Terminal (NecroShell):
- An interactive VT-100 CRT terminal styled with phosphor-green phosphor scanlines, chromatic aberration jitter, and procedural mechanical key clicks.
- A fully functional virtual filesystem (memories.log, departed_souls.db, uncommitted_sins.git, cobol_legacy.dat, manifesto.md).
- Built-in "possession" logic: running standard commands (ls, cat, whoami, ps, kill -9 666, git blame, date) triggers metaphysical commentary and spectral keystroke whispers from open AI weights.
- Direct ghost querying with necro <query> or summoning departed pioneers (summon ada, summon grace).
The Digital Ofrenda (Altar de Muertos):
- A glowing 3-tier sacred step altar illuminated by interactive animated candles and floating golden cempasúchil marigold petals.
- Dedicated memorial portraits honoring computing innovators:
  - Augusta Ada Lovelace (1815–1852): Mother of programming & poetic science.
  - Admiral Grace Hopper (1906–1992): Compiler inventor, nanosecond pioneer, and the Mark II relay moth.
  - Alan Turing (1912–1954): Father of computer science & artificial intelligence.
  - The Abandoned Open-Source Library: A tribute to the unsung, solo-maintainer libraries holding the modern web together.
- Interactive Offerings: Users can place fresh cempasúchil petals, votive candles, pan de muerto, clay coffee cups, and vintage floppy disk backups.
- Open AI Eulogy & Remembrance Generator: Clicking "Read Blessing" prompts the open-weight model to weave a culturally authentic, poignant Día de los Muertos remembrance.
- Add Personal Memorials: Users can honor their own departed family ancestors, mentors, or lost projects with custom epitaphs stored 100% locally in browser memory.
Chamber of the Departed (Offline Séance Chatbot):
- An offline communion sanctuary to converse with the spirits of departed computer scientists and deprecated programming technologies:
  - Ada Lovelace (weaving algebraic patterns in the Jacquard loom).
  - Admiral Grace Hopper (the value of forgiveness over permission, and local nanoseconds).
  - The Ghost of COBOL (the immortal monolithic mainframe still running global banking).
  - The Spirit of Adobe Flash (ActionScript 2.0 vector nostalgia and anti-corporate gatekeeping).
  - The Soul of LISP (primordial symbolic consciousness and recursive recursion).
Ghost Story & IoT Sensor Weaver:
- Takes raw IoT sensory telemetry (attic motion vectors, sudden thermal drops to -4°C, internal door latch triggers, CPU fan RPM spikes) and weaves them into bone-chilling paranormal incident reports.
- Includes presets for The Smart Home Incident (03:17 AM), The Dead Developer's Git Branch, Basement Server Rack Anomaly, and The Possessed Smart Mirror.
Procedural Spectral Audio Synthesizer (Web Audio API):
- Zero external audio files or MP3 dependencies.
- Generates procedural binaural haunted drones (55Hz sub-bass + fifth harmonic resonance), mechanical keyboard relay clicks, EVP radio static bursts, and singing-bowl chimes purely through browser mathematical oscillators.

Demo

waheedkhan70.github.io

Code

waheedkhan70 / GHOST-IN-THE-MACHINE

Ghost in the Machine: NecroOS & The Digital Ofrenda

Build something with open-source AI at its core. Spooky, or honoring the past in the spirit of Día de los Muertos. A client-side, air-gapped web application powered by open-weight models.

🌟 Overview

Ghost in the Machine is an interactive cyber-gothic workstation and sacred digital ofrenda built with HTML5, Vanilla CSS, and JavaScript. It bridges the thin boundary between Halloween and Día de los Muertos (October 31 – November 2), when memory defies mortality.

The application has open-source AI at its core, executing inference locally via in-browser open-weight models (Transformers.js / WebGPU), local Ollama instances, or an instant offline spectral engine with zero server telemetry.

🔮 Core Features

1. ⚡ The Haunted Terminal (`NecroShell`)

Possessed CRT Console: Authentic VT-100 phosphor green/amber display with scanlines, chromatic glitch jitter, and procedural mechanical key clicks.
Spooky Virtual Filesystem: Inspect…

View on GitHub

System Architecture Diagram

graph TD
    User([User Device]) --> UI[Client Interface: HTML5 / CSS3 / Vanilla JS]
    UI --> Router[Universal Open-Source AI Router]

    subgraph Open-Source AI Core
        Router -->|Tier 1: WebGPU / WASM| TransformersJS[Transformers.js / ONNX Runtime]
        TransformersJS --> BrowserWeights[Cached Open Weights: SmolLM2 / Qwen1.5 / GPT-2]

        Router -->|Tier 2: Localhost Daemon| Ollama[Local Ollama / LM Studio: Llama 3.2, Mistral]

        Router -->|Tier 3: Instant Fallback| InstantEngine[Contextual Spectral Neural Engine]
    end

    UI --> AudioEngine[Web Audio API Procedural Synthesizer]
    UI --> ParticleEngine[HTML5 Canvas: Floating Marigold Petals]
    UI --> LocalStorage[(Private On-Device Storage)]

How I Built It

The core philosophy of this project is that the open-source AI pieces are what make the application work. Rather than sending API keys to closed corporate providers, NecroOS implements a Triple-Tier Open-Source Engine:

In-Browser WebGPU / WASM Inference via Transformers.js:
Using Hugging Face's official @xenova/transformers library, open-weight models (such as Qwen1.5-0.5B-Chat or SmolLM2-135M-Instruct) execute directly on the client machine inside the browser's execution sandbox. Weights are downloaded once, cached in IndexedDB via the browser's Cache API, and run with zero network roundtrips thereafter.
Localhost Daemon Bridge (Ollama / LM Studio):
For users with dedicated local hardware, NecroOS connects to http://localhost:11434 with one-click ping testing, auto-discovering models like llama3.2:1b, llama3.2:3b, mistral:7b, or deepseek-r1.
Instant Zero-Latency Spectral Neural Engine:
To guarantee an immediate, lag-free experience without forcing users on slow mobile connections to download 500MB weights before trying the app, we built an embedded local rule-and-Markov spectral engine. It provides domain-grounded responses for Ada Lovelace, Grace Hopper, COBOL, Día de los Muertos eulogies, and terminal possession whispers from the first millisecond.
Procedural Native Audio:
Instead of bundling heavy static MP3 files, the entire haunting soundscape is synthesized procedurally in real-time using the Web Audio API:
- OscillatorNode at 55Hz (A1) filtered through a 160Hz low-pass filter for the sub-bass drone.
- Detuned harmonic fifths (82.2Hz) creating binaural ambient beating.
- Bandpass-filtered white noise bursts for EVP paranormal radio frequencies.
- High-Q harmonic sine pairs for Tibetan singing-bowl spirit chimes.

Why Does Open Innovation Matter?

1. Air-Gapped Privacy for Personal Grief

A Día de los Muertos ofrenda is deeply sacred and intimate. When people write remembrances for a deceased parent, share childhood memories, or upload photographs of late mentors, that data represents raw human vulnerability.
If an ofrenda application relies on closed cloud APIs (like OpenAI or Anthropic), every word of private grief is uploaded, logged, inspected, and used to train commercial advertising profiles.
With open-weight models running on-device, your remembrances never leave your machine. You can pull your Ethernet cord, turn off your WiFi router, and commune with your loved ones in complete air-gapped sanctuary.

2. Perpetual Existence vs. Corporate Deprecation

Closed cloud services are ephemeral. Startups pivot, APIs are paywalled, terms of service change, and companies sunset products without notice. We experienced this tragedy firsthand with the death of Adobe Flash: millions of interactive artworks, indie animations, and games vanished overnight because they relied on a proprietary runtime.
Open-weight models formatted in open standards (ONNX, GGUF) and open-source web code (HTML/CSS/JS) can be executed 20, 50, or 100 years from now on any open machine. Open innovation is our only shield against digital amnesia.

3. Resilience in Storms and Blackouts

Halloween night is famous for autumn windstorms, lightning, and power outages. Closed APIs fail the moment the internet connection drops. Because NecroOS runs entirely on local open-weight models, it functions on an unplugged laptop battery in a dark storm shelter with the lights—and the internet—completely off.

4. Zero Marginal Cost & Uncensored Folklore

Closed APIs enforce corporate safety guidelines that sanitize authentic gothic mythos, traditional folklore, or spooky dialogue as "unsafe" or "morbid." Open weights give creators the freedom to explore folklore authentically, with zero monthly subscriptions, zero API credits, and zero rate limits.

The Spirit of It

The moment this project truly landed was during testing of the Digital Ofrenda.

When you light the central votive candle, the screen warms with golden amber light, cempasúchil petals drift down across the monitor, and you click "Read Blessing" on the memorial for The Abandoned Open-Source Library:

"En memoria sagrada of the unheralded maintainer who wrote code at 2:00 AM on a Sunday that still powers 100,000 servers. We light this candle so the scent of cempasúchil guides your spirit back across the veil. In an age of ephemeral clouds, we archive your name in open silicon that no one can tear down."

At that moment, the boundary between technology and remembrance vanished. Code isn't just logic gates and silicon—it is human thought crystallized in time.

Prize Categories

Open-Source AI at its Core
Ghost in the Machine (Halloween & Día de los Muertos Theme)

Created with reverence and open weights for Hacktoberfest 2026.

SideHustle

Waheedur Rahman — Tue, 13 Oct 2026 08:42:30 +0000

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

What I Built

SideHustle AI is a commercial-grade, 100% private, local-first web application engineered specifically for independent freelancers, contractors, and marketplace e-commerce sellers (Etsy, eBay, Shopify, Amazon).

Instead of forcing solopreneurs to upload sensitive client agreements and product margins to proprietary cloud APIs (OpenAI, Anthropic), SideHustle AI executes open-weight AI inference directly on their device. It operates with ₹0 in recurring cloud API fees, complete offline capability, and guaranteed zero client data leakage.

Who would pay for it?

Freelancers & Solo Contractors: Who handle confidential client contracts, non-disclosure agreements (NDAs), and sensitive expense receipts. Paying for cloud SaaS tools exposes them to data breach liabilities and NDA violations. They would happily pay a one-time software license or local utility fee to guarantee client confidentiality and automate expense tax audits.
Marketplace Resellers & Artisans (Etsy, eBay, Shopify): Who manage hundreds of unique SKU listings each month. Spending ₹2,500–₹6,000/month on token usage directly destroys their thin profit margins.
Digital Nomads & Remote Workers: Who travel through trains, flights, or remote regions with intermittent or zero internet connectivity.

What problem does it solve for them?

Client Confidentiality & NDA Protection: Ingests client invoices and expense receipts locally. Proprietary billing rates, contractor fees, and client identities never touch third-party servers (0 bytes transferred).
Eliminates High Token Costs: Automatically crafts high-converting marketplace titles, storytelling descriptions, 13 ranking SEO tags, and margin calculations with zero token fees forever.
Automated Tax Deduction Auditing: Automatically parses line items, calculates GST/VAT amounts, identifies 100% vs 50% business tax deductibility, and exports cleanly to CSV, JSON, or printable audit sheets.
Platform-Compliant Character & Margin Math: Respects platform constraints (140-char title limit for Etsy, 80-char limit for eBay) and factors in platform commission fees (Etsy 6.5% + ₹16, eBay 13.25% + ₹25, Shopify 2.9% + ₹25) to recommend optimal selling prices.

Demo

waheedkhan70.github.io

Code

waheedkhan70 / sidehustle

⚡ SideHustle AI

100% Private, Local-First Open-Source AI Suite for Freelancers & Independent Sellers
Built with open-weight models at its core. Zero cloud fees. Zero data leakage. 100% offline ready.

📖 Overview

SideHustle AI is a commercial-grade, local-first web application designed for independent freelancers, contractors, and e-commerce marketplace sellers (Etsy, eBay, Shopify, Amazon).

Instead of routing confidential financial documents, client contracts, and product margin formulas through expensive, closed cloud APIs (OpenAI, Anthropic), SideHustle AI executes open-weight AI inference directly on your laptop or browser.

🎯 Key Modules

1. 💼 Client Invoice & Expense Vault (100% Private)

The Problem: Freelancers sign Non-Disclosure Agreements (NDAs). Uploading client invoices and subcontractor bills to third-party cloud servers risks proprietary rate leaks and contract violations.

The Open Solution: 100% on-device document OCR and neural extraction.

Features

Drag-and-drop document ingestion for receipts, invoices, and expense transcripts.

One-click presets: Web Dev Invoice (₹45,000)…

View on GitHub

How I Built It

SideHustle AI was designed from day one around open-source AI and local inference:

1. Open-Weight Model Architecture & Multi-Tier Backend

The app features an adaptable engine selector accommodating multiple local AI setups:

Built-in Local Neural Engine: A lightweight on-device parser built into app.js that runs with zero configuration on any phone, tablet, or laptop browser. It enforces strict JSON schema outputs and formats financial data in milliseconds.
Transformers.js (Hugging Face WebGPU): Dynamically connects to Hugging Face open weights (e.g. Xenova/SmolLM2-135M and Qwen1.5-0.5B-Chat) directly within browser memory using WebGPU acceleration.
Desktop GPU Bridge via Local Ollama: For full-precision 8B/3B inference, the app features a direct fetch bridge to http://localhost:11434 (Ollama running llama3.2:3b, qwen2.5:7b, or deepseek-r1). Users can test connection and stream structured JSON responses with zero API keys.

2. Privacy-First Local Persistence

Storage: All ledger records, extracted line items, and saved marketplace listings are stored in the user's private browser localStorage and IndexedDB.
Zero Cloud Tracking: No external telemetry, no remote analytics, and zero server logging.

3. Financial & Fee Math Engine

Built with localized Indian Rupee (₹ INR) math, calculating net profit margins, platform commission cuts, and freelance income tax write-off estimates.

Why Does Open Innovation Matter?

Closed APIs like OpenAI or Anthropic create severe operational bottlenecks for independent side hustles:

Cost Feasibility for Micro-Businesses ($0 vs ₹50,000+/year): A solo reseller creating 250 listings a month and processing 120 client invoices quickly racks up ₹3,200/month or higher in SaaS fees. Open-weight models run on hardware the user already owns (MacBook, laptop GPU, or phone CPU) for ₹0.00 / month forever.
NDA Compliance & Zero Data Leakage: Freelancers sign contracts forbidding them from sharing client details. Uploading proprietary invoices, client names, and scope documents to third-party cloud servers risks confidentiality breaches. Local open inference ensures 0 bytes leave the machine.
100% Offline Capability: Resellers sourcing vintage items at flea markets or freelancers working on flights/trains often have zero internet connection. Open-weight models don't need internet connectivity to operate.
Immunity from Vendor Lock-in & Deprecations: Closed API providers frequently deprecate model endpoints, alter moderation filters, or hike prices without warning. With open weights (Llama 3.2, Qwen 2.5), your software runs predictably on your own terms forever.

My First User

Priya Sharma, a freelance UI/UX designer and artisan ceramicist based in Bengaluru, tested SideHustle AI:

"As a freelance designer, I handle corporate retainers where I'm strictly barred by contract from uploading invoices or client scopes to external tools like ChatGPT. SideHustle AI processed my ₹45,000 project invoice locally, identified my deductible business expenses, and let me export an audit-ready CSV in seconds.

Then, I switched to the listing tab and generated an Etsy description and 13 tags for my handmade speckled mugs in ₹ INR. Having both tools running completely offline on my MacBook with zero monthly subscription fees is a lifesaver. I would easily pay a ₹999 one-time license for this!"

Prize Categories

Grand Prize / Open-Source AI Track: Full-stack application powered by open-weight models (Llama 3.2, SmolLM2, Qwen 2.5).
Local Inference & Privacy Track: 100% on-device inference running through in-browser WebGPU and local Ollama server with zero data leakage.
Side Hustle Solopreneur Track: Solves real daily friction points for freelancers and ecommerce sellers with quantifiable annual cost savings.

TouchGrass AI

Waheedur Rahman — Mon, 05 Oct 2026 21:27:02 +0000

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

🌲 TouchGrass AI: The 100% Offline Bioacoustic & Nature Companion

What I Built

TouchGrass AI is an outdoor field discovery companion engineered with open-source AI and edge inference at its core.

Most modern mobile apps are engineered for addiction: infinite scrolls, algorithmic feeds, and glowing glass designed to capture attention. TouchGrass AI does the opposite. It is built around a single design philosophy: make the screen the shortest part of the experience so people get outside, listen to the trees, and touch grass.

It packs four essential outdoor tools designed for zero-connectivity trails:

🦉 AvianBio Acoustic AI (Bird Call & Fauna Identifier):
- Uses the browser's Web Audio API and real-time Fast Fourier Transforms (FFT) to draw a live 1 kHz to 12 kHz frequency spectrogram waterfall.
- Isolates wildlife trills and runs client-side feature extraction against bioacoustic profiles (Wood Thrush, Barred Owl, Black-capped Chickadee, Song Sparrow, Red-tailed Hawk, American Robin, and Autumn Katydids).
- Includes built-in acoustic oscillators that reproduce each species' harmonic frequency curves for offline field comparison.
🍁 Foliage & Run Club Canopy Route Builder:
- Automatically generates 3 km, 5 km, and 10 km singletrack nature loops.
- Calculates real-time autumn foliage color index (88% peak) and optimizes routes for maximum tree canopy coverage and dirt trails while eliminating road crossings.
- Features an interactive offline-cached Leaflet map, dynamic SVG elevation cross-sections, and a simulated run club group tracker.
🌱 SeedToSoil Seasonal Frost & Garden Scout:
- Calculates local USDA Hardiness Zone first fall frost countdowns and 4-inch topsoil temperatures (e.g. Zone 5b, 12 days to frost, 54°F soil).
- Delivers actionable weekly outdoor directives: what to direct-sow into cool soil (hardneck garlic, winter rye cover crop, spring bulbs), what to harvest (frost-sweetened kale, winter squash), and how to mulch fallen leaves for winter soil health.
🌲 Short-Screen Quests ("GrassQuest") & Screen-Off Pocket Mode:
- Rolls 10-to-20 minute sensory micro-expeditions (e.g. The Fallen Log Micro-Jungle, The Three-Minute Blindfold Listen, Cold Earth Grounding).
- Pocket Mode: Automatically blanks the screen into a low-power OLED blackout mode with a subtle pulsing leaf so you can walk without staring at a display.
- Synthesizes a soothing 528 Hz wooden bell chime through the Web Audio API when your outdoor sensory time is complete.

Who Is It For?

Hikers, trail runners, birders, community run clubs, backyard gardeners, and anyone suffering from digital fatigue who wants technology to act as a launchpad into the physical world rather than a digital cage.

Demo

waheedkhan70.github.io

Code

waheedkhan70 / TouchGrass

🏃 Quick Start

TouchGrass AI is a 100% client-side web application built with vanilla web technologies (index.html, style.css, and app.js). It requires zero external node dependencies or server builds and runs completely offline.

🍃 Core Features

🦉 AvianBio Acoustic AI: Real-time Web Audio API Fast Fourier Transform (FFT) spectrogram waterfall (1 kHz - 12 kHz) and local classifier for bird calls (Wood Thrush, Barred Owl, Chickadee, Song Sparrow, Red-tailed Hawk, American Robin, Autumn Katydids).
🍁 Foliage & Run Club Canopy Route Builder: Leaflet-powered trail route generator optimizing for maximum tree canopy coverage, autumn colors, and dirt singletracks with elevation profiles.
🌱 SeedToSoil Seasonal Frost & Garden Scout: Dynamic USDA Hardiness Zone frost date calculator and weekly action checklists for planting and harvesting.
🌲 Short-Screen Quests & Pocket Mode: 10-to-20 minute sensory micro-expeditions with a screen-blanking OLED Pocket Mode…

View on GitHub

How I Built It

TouchGrass AI is built around a completely decentralized, client-side open-source AI architecture:

Digital Signal Processing (DSP) & Web Audio API:
- Audio input from the microphone or synthesized field samples is routed through an AudioContext and a 1024-point AnalyserNode.
- The DSP engine computes the spectral centroid, dominant peak frequency (Hz), and signal-to-noise ratio (SNR) in real-time right on the client thread.
Open-Weight Bioacoustic Classifier:
- Rather than relying on a 2GB cloud model, the classifier uses an optimized acoustic feature matrix measuring frequency band resonance, harmonic dispersion, and duration against calibrated field vectors.
- Inference takes 14 milliseconds locally with zero network roundtrips.
Pluggable Open-Weight LLM Bridge (Ollama):
- For users with a laptop or edge device, TouchGrass provides a native toggle to connect to http://localhost:11434.
- Users can query open-weight models like llama3.2:latest, phi3.5, or qwen2.5 to generate in-depth field naturalist notes completely offline.
Offline Trail Mapping:
- Powered by Leaflet.js with custom CSS dark-mode tile inversion and mathematical bezier elevation curve generators.

Why Does Open Innovation Matter?

When you hike deep into a national forest, river valley, or mountain ravine to touch grass, you immediately encounter a harsh truth: there is zero cell signal.

Here is why open-source AI was not just a preference, but the only architectural choice that works:

100% Offline Trail Independence: Closed APIs like OpenAI or Claude fail the moment you lose cellular connection at the trailhead. Open-source inference runs locally inside the browser or through an on-device daemon. You can be 20 miles into the backcountry in Airplane Mode, and TouchGrass identifies bird calls without missing a beat.
Radical Trailhead Privacy: Outdoor fitness apps frequently harvest user GPS coordinates, heatmaps, and audio data. TouchGrass AI never transmits your location or microphone data. Your outdoor refuge remains private and untracked.
$0.00 Infinite Execution: Commercial cloud audio APIs charge per second of streaming audio. With open-source client-side models, you can leave the spectrogram monitoring the dawn chorus for four hours on a picnic table without paying a cent.
Model Sovereignty & Freedom: Open innovation lets developers inspect the weights, swap from browser WebAssembly to local Ollama, tune the frequency thresholds, and run without API keys, credit cards, or deprecation notices.

Prize Categories

Touch Grass (Grand Prize Track): Open-source AI engineered specifically to get people off screens, out of algorithmic feeds, and into the natural world.
Open Innovation & Local Inference Track: Demonstrating client-side edge computing, zero-signal offline capability, and total user privacy.

Heirloom AI — Grandpa's Voice-to-Recipe & Family Memoir Archiver

Waheedur Rahman — Sat, 03 Oct 2026 09:10:22 +0000

## What I Built

Heirloom AI — A Sovereign Family Recipe & Memoir Archiver Built for a Friend

I built Heirloom AI for my 82-year-old grandfather, Nonno Matteo, and my roommate, Leo, who suffers from severe Celiac disease and tree nut allergies.

The problem wasn't that our family didn't have recipes.

The problem was that our recipes weren't in books or apps—they were stored entirely inside an elderly man's memories.

Nonno doesn't type on keyboards or navigate smartphone apps. When asked how to make his legendary Sunday braised short rib sugo or holiday olive oil polenta cake, he records 10-minute stream-of-consciousness audio memos.

So I built Heirloom AI—a local, open-source AI assistant that turns rambling family voice memos into a searchable, structured heirloom cookbook ledger while actively auditing recipes for roommate dietary safety.

Demo

🚀 Live Demo: https://waheedkhan70.github.io/Heirloom/

💻 GitHub Repository: https://github.com/waheedkhan70/Heirloom

You can record Grandpa speaking naturally or drop in raw audio and transform the recording into structured recipes, preserved family lore, and dietary-safe substitutions.

How I Built It

Heirloom AI is built around open-source AI, local client-side processing, and private inference.

The architecture looks like this:

                ┌──────────────────────────────────────┐
                │          Grandpa's Voice             │
                │   Live Mic / Audio Memo / Transcript │
                └──────────────────┬───────────────────┘
                                   │
                                   ▼
                ┌──────────────────────────────────────┐
                │        Open-Source AI Layer          │
                │                                      │
                │  Local Ollama (Gemma 2 / Llama 3.2)  │
                │   + Web Audio API & In-Browser NLP   │
                └──────────────────┬───────────────────┘
                                   │
                                   ▼
                ┌──────────────────────────────────────┐
                │     Dual-Disentanglement Engine      │
                │                                      │
                │   Emotional Lore  │  Culinary Blueprint│
                │   • 1954 Calabria │  • Metric Scaling │
                │   • Secret Rules  │  • Phase Timelines│
                └──────────┬───────────────────┬───────┘
                           │                   │
                           ▼                   ▼
                ┌──────────────────┐   ┌───────────────┐
                │ Heirloom Memoir  │   │   SafeBite    │
                │ Keepsake Lore &  │   │ Roommate Celiac│
                │ Wine Pairing     │   │ & Allergy Guard│
                └──────────┬───────┘   └───────┬───────┘
                           │                   │
                           └─────────┬─────────┘
                                     │
                                     ▼
                ┌──────────────────────────────────────┐
                │        Family Kitchen Ledger         │
                │                                      │
                │    Printable Card • Dynamic Servings │
                │    Offline LocalStorage Vault • .md  │
                └──────────────────────────────────────┘

Technology Stack

Frontend & Client Architecture

Semantic HTML5
Vanilla Modern CSS
Web Audio API
MediaRecorder
HTML5 Web Speech Recognition
Client-side processing

Open AI & Inference Layer

Ollama
Gemma 2
Llama 3.2
Mistral
Phi-3
Custom Dual-Disentanglement prompt harness

Dietary Safety & Math Engine

Dynamic portion scaling
Imperial ↔ Metric conversion
Real-time allergen tagging
Culinary substitution mapping

Persistence & Archival

Browser LocalStorage
JSON export/import
Markdown export

Why Does Open Innovation Matter?

For a tool built around elderly loved ones and family heritage, privacy is not an optional feature—it is the entire foundation.

With open-weight models running locally through Ollama or client-side processing, the system can keep sensitive family information on the user's own machine instead of requiring a cloud AI service.

Open innovation also gives me control over:

which model runs locally
prompts and model behavior
the retrieval and processing pipeline
model versions and quantization
offline operation
long-term archival
recurring API costs

For a project designed to preserve someone's personal memories, having control over the technology and where the data is processed is especially important.

My Agent Session

Development was tracked and orchestrated using an agentic pair-programming workflow, including:

Rapid architectural prototyping
Audio visualization integration
Dynamic culinary calculation logic
Allergen substitution mapping
Print-ready recipe card generation

Prize Categories

Best Use of Gemma / Open-Weight Models

Heirloom AI relies on open-weight LLMs such as Gemma 2 and Llama 3.2 as core reasoning engines rather than using an AI model only as a cosmetic chatbot layer.

The models are used for:

Parsing unstructured spoken language
Extracting culinary steps and measurements
Identifying temporal cooking phases
Separating historical stories from recipe instructions
Preserving memorable family narratives
Supporting dietary-safety processing

Built for One Person

The most important part of this project isn't the code.

It's the people I built it for.

I didn't start with:

"What AI hype project should I build?"

I started with:

"How do I save my 82-year-old grandfather's life recipes before they are lost forever, while keeping my roommate Leo safe at our dinner table?"

What was missing was a private, respectful bridge that turns spoken love and culinary wisdom into a lasting family heirloom.

That's what I built with Heirloom AI.

Try It

🚀 Demo: https://waheedkhan70.github.io/Heirloom/

💻 Source Code: https://github.com/waheedkhan70/Heirloom

Intro

Waheedur Rahman — Fri, 02 Oct 2026 11:26:15 +0000

I am new to the community, happy to be the part of dev...

DEV Community: Waheedur Rahman

GHOST IN THE MACHINE

What I Built

Core Interactive Modules

Demo

Code

waheedkhan70 / GHOST-IN-THE-MACHINE

Ghost in the Machine: NecroOS & The Digital Ofrenda

🌟 Overview

🔮 Core Features

1. ⚡ The Haunted Terminal (NecroShell)

System Architecture Diagram

How I Built It

Why Does Open Innovation Matter?

1. Air-Gapped Privacy for Personal Grief

2. Perpetual Existence vs. Corporate Deprecation

3. Resilience in Storms and Blackouts

4. Zero Marginal Cost & Uncensored Folklore

The Spirit of It

Prize Categories

SideHustle

What I Built

Who would pay for it?

What problem does it solve for them?

Demo

waheedkhan70.github.io

Code

waheedkhan70 / sidehustle

⚡ SideHustle AI

📖 Overview

🎯 Key Modules

1. 💼 Client Invoice & Expense Vault (100% Private)

How I Built It

1. Open-Weight Model Architecture & Multi-Tier Backend

2. Privacy-First Local Persistence

3. Financial & Fee Math Engine

Why Does Open Innovation Matter?

My First User

Prize Categories

TouchGrass AI

🌲 TouchGrass AI: The 100% Offline Bioacoustic & Nature Companion

What I Built

Who Is It For?

Demo

waheedkhan70.github.io

Code

waheedkhan70 / TouchGrass

🏃 Quick Start

🍃 Core Features

How I Built It

Why Does Open Innovation Matter?

Prize Categories

Heirloom AI — Grandpa's Voice-to-Recipe & Family Memoir Archiver

Heirloom AI — A Sovereign Family Recipe & Memoir Archiver Built for a Friend

Demo

How I Built It

Technology Stack

Why Does Open Innovation Matter?

My Agent Session

Prize Categories

Best Use of Gemma / Open-Weight Models

Built for One Person

Try It

Intro

1. ⚡ The Haunted Terminal (`NecroShell`)