DEV Community: Anshuman Singh

🏵️ Ofrenda Viva: Honoring the Ghosts in Our Machines with 100% Local Open-Source AI

Anshuman Singh — Sat, 31 Oct 2026 14:48:39 +0000

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

What I Built

"There are three deaths in this world:

The first is when our body ceases to breathe.

The second is when they lower us into the earth.

And the third, the true and final death, is when no one speaks our name again."

— Traditional Mexican Adage

Between the twilight of Halloween on October 31 and the sacred dawn of Día de los Muertos (Day of the Dead) on November 1 and 2, ancient folklore tells us that the veil separating the living from the departed wears thin. Across Mexico and throughout homes worldwide, families assemble ofrendas—multi-tiered altars illuminated by flickering candlelight (velas), draped in golden marigolds (cempasúchil), and adorned with favorite foods, portraits, and memories of those who have crossed over.

For the "Ghost in the Machine" challenge, I built Ofrenda Viva ("Living Offering"): an interactive, audio-visual digital memorial sanctuary and local AI spirit oracle. It allows anyone to pay homage to the legendary computing pioneers whose ideas literally live inside our hardware and software, or build an air-gapped memorial altar for their own departed family members, mentors, and beloved pets.

Core Features:

🏛️ The Three-Tiered Altar (Los Tres Niveles): Authentically structured into El Cielo (The Celestial Realm with the centerpiece portrait shrine), La Tierra (Earthly Life with handcrafted Sugar Skulls and Pan de Muerto), and El Inframundo (Purification with salt, spring water, and smoldering copal incense).
🕯️ Interactive Candlelit Rituals: Light 6 dedicated veladoras (candles) with custom procedural sound synthesis—each illuminating the altar with warm realistic radial glows.
💾 "The Ghosts in Our Machines": Dedicated altars for computing pioneers—Ada Lovelace, Alan Turing, Grace Hopper, and Dennis Ritchie—whose open innovations laid the foundations of modern computing.
🕊️ Custom Ancestor & Pet Memorials: Add your own ancestors, grandparents, departed mentors, or family pets with complete offline privacy and photo storage.
📜 "Espíritu" — In-Browser Local AI Oracle: An open-source AI engine that synthesizes Epitafios Eternos (poetic eulogies), Calaveritas Literarias (traditional rhyming octosyllabic Mexican folk verses), Cartas al Más Allá (two-way séance letters), and Guías de Ofrenda (ritual offering recommendations) with real-time typewriter streaming.
🎸 Pure Procedural Web Audio Soundscape: Zero audio MP3 downloads—synthesizes Spanish classical guitar arpeggios, Tibetan singing bowl harmonic chimes, and fire crackling using mathematical oscillators in the native Web Audio API.
🏵️ 60 FPS Marigold & Copal Canvas Engine: A high-performance physics particle engine simulating drifting golden cempasúchil petals, interactive wind mouse deflection, and rising copal smoke wisps.

Demo

🎮 Live Web Application: https://anshumansingh0408-hub.github.io/hbs/ (or clone & run locally via npm run dev)
🎬 Full Walkthrough Video: Included in the repository as ofrenda_viva_demo.mp4 (H.264, 1366x644, 29.7s walkthrough showing altar lighting, pioneer switches, and AI token streaming).
🖥️ Standalone Local Video Player: Open play_demo.html in any browser to watch the demo video with custom playback controls and direct downloads.

Interactive Highlights

Landing Hero & Bento Showcase: Modern baroque aesthetics with glowing typography (Cinzel Decorative & Outfit) and quick preview cards.
Interactive Altar Stage: Click on any of the 6 candles or the copal incense bowl to trigger harmonic chimes, flame animations, and cultural folklore lore panels.
AI Oracle Console: Select a spirit, choose a literary mode, and watch tokens stream into view with rhythmic punctuation cadence.

Code

All code is open-source under the permissive MIT License:

anshumansingh0408-hub / hbs

🏵️ Ofrenda Viva: Digital Sanctuary of Remembrance & Local Open-Source AI Oracle

Ghost in the Machine — Día de los Muertos & Open-Source AI Challenge Submission
A 100% private, air-gapped memorial altar powered by client-side open-weight models, procedural Web Audio synthesis, and real-time marigold petal physics.

🌟 Overview

Ofrenda Viva ("Living Offering") is a digital sanctuary honoring Día de los Muertos (Day of the Dead) and Halloween, built on a fundamental principle: Grief, ancestral memory, and intimate human eulogies belong solely to the individual, never to centralized corporate surveillance clouds.

In traditional Mexican heritage, an ofrenda is a sacred three-tiered altar created to welcome back the souls of loved ones on November 1st and 2nd. The veil between past and present thins, guided by glowing candles (velas), the golden scent of marigolds (cempasúchil), fragrant copal incense, and heartfelt memories.

Ofrenda Viva brings this tradition into the…

View on GitHub

Repository: https://github.com/anshumansingh0408-hub/hbs

Architecture Overview


mermaid
graph TD
    User([User / Memory Keeper]) --> UI[Sanctuary UI: Tiered Altar & Bento Grid]

    subgraph Client_Sanctuary [100% Local In-Browser Execution]
        UI --> PetalsEngine[Canvas Particle Engine: Cempasúchil Petals & Copal Smoke]
        UI --> WebAudio[Procedural Web Audio API: Spanish Guitar, Chimes & Fire Crackle]
        UI --> AltarState[Altar State Manager: 6 Interactive Candles & Offerings]
        UI --> Storage[(Private LocalStorage: Custom Photos & Memorials)]

        subgraph Open_AI_Core [Espíritu Local AI Engine]
            AltarState --> AIOrchestrator[Folkloric Prompt & Personality Harness]
            Storage --> AIOrchestrator
            AIOrchestrator --> LocalInference[Local Open-Weight Model / Client Pipeline]
            LocalInference --> StreamingPipeline[Token-by-Token Streaming Engine]
        end
    end

    StreamingPipeline --> UI
    style Open_AI_Core fill:#2b1810,stroke:#ffb703,stroke-width:2px;
    style Client_Sanctuary fill:#0d0914,stroke:#7209b7,stroke-width:2px;

SlothOps: Channeling Your Inner Sloth with Open-Weight AI & TabPFN-3.5

Anshuman Singh — Sun, 25 Oct 2026 17:23:07 +0000

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

What I Built

SlothOps is an autonomous, privacy-first digital chore autopilot built to celebrate International Sloth Day and reward human laziness.

Sloths move at their own pace, conserve energy, and do only what matters. Knowledge workers and developers, on the other hand, spend hours every week on mundane digital chores they dread:

Sorting chaotic Downloads/ and Desktop/ folders with messy screenshots and invoices.
Manually auditing banking statements or CSV exports to catch sneaky subscription price creep.
Wading through dozens of unread newsletters, Discord pings, and GitHub notifications.
Clicking through dozens of tabs for routine research.

SlothOps eliminates all of them. In a single click, it executes the entire digital maintenance cycle:

Sunday Clutter Sweep: Google Gemma semantically classifies and files messy downloads into organized project directories.
Subscription Vampire Slayer: Prior Labs TabPFN-3.5 foundation model performs zero-shot in-context learning over banking transactions to flag 20%+ price spikes and zombie subscriptions.
Morning Zero-Effort Digest: Condenses 100+ feeds into 2 actionable bullets.
Sloth Audio Lounge: Synthesizes an ambient ElevenLabs voice narration so you can literally close your eyes, lie in bed, and listen to the day's briefing with zero screen glare.

Demo

Live Interactive Web App: Hosted on Render(https://slothops-agent-w8f1.onrender.com)
GitHub Repository: https://github.com/anshumansingh0408-hub/slothops[](url)
Local Air-Gapped Mode: Runs 100% offline via local Gemma inference and TabPFN in-context forward passes.

Code

The full source code is licensed under the permissive MIT License and structured as follows:

slothops/
├── server/
│   ├── gemma_engine.js           # Google Gemma open-weight runtime (Ollama/vLLM/Local)
│   ├── mastra_orchestrator.js    # Mastra DAG engine with Temporal-style retries
│   ├── observability.js          # Sentry Agent Tracing span recorder & token accountant
│   └── tools/
│       ├── tabpfn_tool.js        # Prior Labs TabPFN-3.5 zero-shot tabular AI
│       ├── file_tidier_tool.js   # Automated local folder declutter
│       ├── inbox_condenser_tool.js # 2-bullet morning digest
│       ├── browser_agent_tool.js # Browserbase Stagehand autonomous web researcher
│       ├── voice_digest_tool.js  # ElevenLabs calming audio synthesizer
│       └── memory_vector_tool.js # Tiger Data pgvector / MongoDB Atlas vector memory
└── src/                          # Modern glassmorphic React + Vite user interface

How I Built It

SlothOps is designed strictly around open-source AI at its core:

Google Gemma 2 (Open-Weights): The central reasoning brain. It parses complex file contexts, plans multi-step chore execution, and handles structured JSON output generation without proprietary API lock-in.
Mastra Framework & Temporal Durability: Workflows are orchestrated using a TypeScript DAG pipeline with state machine durability, step idempotency, and exponential backoff retry policies.
Prior Labs TabPFN-3.5 (Tabular Foundation Model): Rather than writing manual regexes or training complex XGBoost pipelines with cross-validation, TabPFN-3.5 ingests banking transactions in a single forward pass, predicting end-of-month cash burn and flagging anomalous subscription creep.
Browserbase (Stagehand SDK): Enables the agent to navigate web documentation and partner portals autonomously.
Tiger Data (PostgreSQL + pgvector) & Sentry: Stores semantic user preferences across sessions and tracks agent execution spans, latency, and token consumption.

Why Does Open Innovation Matter?

Why did an open-source, open-weight approach work significantly better than a closed proprietary API?

Total Financial and Personal Privacy: Bank transactions, tax receipts, and personal chat messages should never be transmitted to third-party cloud LLM servers. Because Gemma and TabPFN run locally on the machine, your sensitive financial data remains completely air-gapped.
Zero Recurring API Tax: Daily automation should not cost $0.05 every time you clean a folder. With open-weights, running your daily automation costs exactly $0.00. Set it up once, and it runs forever.
100% Offline Resilience: You can run SlothOps on a flight with Wi-Fi disabled or in off-grid environments.
Model Swappability & Fine-Tuning: By decoupling the harness from closed providers, developers can easily swap between Gemma 2B, Gemma 9B, or custom fine-tunes created with Thinking Machines' Tinker.

My Agent Session

I developed and refined SlothOps collaboratively with an autonomous AI coding assistant. The agent assisted in scaffolding the Mastra DAG architecture, implementing TabPFN zero-shot tabular evaluation, styling the luxury glassmorphism UI, and rendering the custom Remotion demo video.

All code, architectural blueprints, and workflows are committed and open-sourced on GitHub: github.com/anshumansingh0408-hub/slothops

The Lazy Payoff

How much time and energy does SlothOps actually save?

47 minutes saved daily from reading repetitive newsletters and group chat pings.
$113.99 saved next month by auto-detecting a 52% price creep on developer subscriptions and a zombie gym fee.
18 minutes of manual file dragging saved every Sunday evening.
100% reduction in screen glare by converting summaries into an ambient voice digest.

Prize Categories

We are entering the following Hacktoberfest Week 3 categories:

Best Use of Gemma: Google Gemma 2 (9B open-weights) serves as the primary reasoning and semantic classification engine.
Best Use of TabPFN (Prior Labs): TabPFN-3.5 tabular foundation model is integrated for zero-shot transaction anomaly detection and end-of-month cash burn forecasting.
Best Use of Render: Complete fullstack deployment on Render using the render.yaml infrastructure-as-code blueprint.
Best Use of DigitalOcean: Docker containerized and ready for 1-Click deployment on DigitalOcean GPU Droplets.

hello

Anshuman Singh — Tue, 20 Oct 2026 12:47:24 +0000