DEV Community: Tatvagna Tandel

Ofrenda Local: An Air-Gapped Memory Altar Powered by Open-Source AI published: true tags: devchallenge, hf26challenge, opensource, ai

Tatvagna Tandel — Sat, 31 Oct 2026 09:20:03 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine (https://dev.to/challenges/hacktoberfest-week4-2026-10-26
)

What I Built

NecroShell is a haunted command-line companion that breathes an eerie, Victorian spirit into your terminal session. It runs silently in the background, listening to standard input/output hooks in your shell (bash/zsh).

Whenever a command fails, a build breaks, a git merge conflicts, or a midnight debugging session drags on, NecroShell intercepts the event and whispers back chilling, gothic commentary, cryptic riddles, and spectral post-mortem diagnostics directly into your terminal.

I interpreted the theme literally: modern computers have millions of forgotten lines of legacy code, zombie processes, and dead threads. NecroShell embodies that digital afterlife as an actual ghost in the machine, giving a voice to the silicon graveyard.
How I Built It

NecroShell was architected from the ground up around open-source AI and air-gapped local inference:

Local Model: Powered by Llama 3.2 (3B) or SmolLM2 (1.7B) quantized to 4-bit (Q4_K_M), small enough to live in laptop RAM without impacting system performance.
Inference Runtime: Powered by Ollama / llama.cpp, ensuring sub-500ms response latencies with zero external network requests.
Shell Hooks: A lightweight POSIX pre-exec and post-exec trap that captures exit codes, terminal errors, and command duration, feeding them as contextual telemetry to the local agent harness.
The Necromancer Prompt Engine: A deterministic system prompt that forces the LLM into character (a sorrowful spirit trapped in silicon) while strictly preserving technical utility (the error diagnosis must be accurate).
Why Does Open Innovation Matter?

This project simply could not exist using closed, proprietary APIs:

True Lights-Out Offline Execution: In the spirit of Halloween, you can take your laptop into an abandoned cabin in the woods, disconnect Wi-Fi completely, and the ghost still whispers back. No internet ping, no server status page, and no cloud outages.
Zero Moderation False Positives: Gothic horror, macabre humor, and eerie descriptions (graveyard, zombie processes, reaping threads) routinely trip the sanitization filters and guardrails of commercial cloud APIs. Open-weight models let developers set authentic, nuanced creative boundaries.
No Terminal Snooping: Your terminal history often contains secret paths, environmental variables, database names, and internal IP addresses. Piping that data to an external API is a security risk. Local inference guarantees your private shell sessions never leave localhost.
Zero Cost per Token: Having an AI process every command exit status would rack up cloud API bills rapidly. With open weights, it costs nothing to run forever.

SlothSnooze: The Air-Gapped Bedside Companion That Rewards Extreme Laziness

Tatvagna Tandel — Thu, 22 Oct 2026 15:27:06 +0000

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

What I Built

Sloths sleep between 15 and 20 hours a day. They don’t doomscroll, they don't count steps, they don't subscribe to wellness apps, and they never stress about tomorrow.

In honor of International Sloth Day and this week’s "Slow and Steady" Hacktoberfest challenge, I built SlothSnooze: an air-gapped, open-source AI bedtime companion designed to reward absolute, unadulterated laziness.

Instead of fighting insomnia by staring into phone screens or paying $70/year to have commercial apps transmit your bedroom audio to corporate clouds, SlothSnooze does all the heavy lifting with a single click:

Generates hyper-soothing, slow-cadence bedtime stories powered by local open-weight models (Llama 3.2, Qwen 2.5, or its built-in offline neural engine) grounded in rainforest sensory details (cool mountain mist, damp moss, twilight canopies).
Synthesizes real-time procedural ambient soundscapes using pure mathematical Web Audio (rain on canopy leaves, gentle wind gusts, night crickets, bubbling mountain brooks, and Tibetan singing bowls).
Entrains your brainwaves into restorative delta/theta sleep using authentic stereo binaural beats (4.0 Hz Theta wave & 2.5 Hz Delta wave).
Narrates the story at an ultra-slow sloth cadence (~75 words per minute) with built-in acoustic ducking—automatically softening ambient rain when the voice speaks and swelling gently during breath pauses.
Slows your pulse with a bioluminescent sloth breathing guide (Sloth 4-7-8 and Deep Hibernation rhythms).
Fades out automatically with an exponential 3-minute sleep timer.

Who gets to be lazy? Anyone who has ever laid awake in bed with a spinning mind. You press one button, roll over, and fall asleep.

Demo

Here is a look at SlothSnooze in action:

Key Features in Action:

1-Click Lazy Sleep Mode: Directly activates Night Dim mode, launches the ambient soundscape, starts the 45-minute canopy snooze timer, and begins slow bedtime narration.
Pure Mathematical Sound Synthesizer: 6-channel mixer running directly on Web Audio API—zero audio files downloaded, 100% offline.
Audio Oscilloscope & Visualizer: Displays live sound frequencies with a bioluminescent emerald glow.
Sloth 4-7-8 Breathing Orb: Gently expands and contracts with glowing firefly particles to guide parasympathetic nervous system relaxation.
Night Dim Mode: Softens blue light to protect melatonin production at your bedside.

Code

The complete source code is available in the repository.

Key Architecture:

Backend (FastAPI + Python):
- app/engine/story_agent.py: Probes local Ollama instances (http://localhost:11434), auto-detects downloaded open weights (Llama 3.2, Qwen 2.5, Mistral), and provides an instant offline fallback engine for completely air-gapped operation.
- app/engine/soundscape_engine.py: Defines acoustic harmonics, binaural beat offsets, and sloth breath timings.
- app/engine/journal_engine.py: Air-gapped local sleep journal that stores dream logs in local JSON files without telemetry.
Frontend (Pure Modern CSS & JavaScript):
- app/static/js/audio_synth.js: Web Audio node network (Pink noise generators, biquad bandpass filters, LFO frequency modulation, stereo channel mergers).
- app/static/js/story_reader.js: Offline Speech Synthesis reader tuned to sloth pacing with dynamic audio ducking.
- app/static/js/app.js: Dynamic starfield canvas, audio oscilloscope, breathing engine, and sleep timer.

How I Built It

SlothSnooze is built from the ground up around open-source AI and local inference:

Open-Weight Language Models:
- SlothSnooze integrates with local open-weight models via Ollama and OpenAI-compatible local servers (like LM Studio or llama.cpp).
- It shines brightest with compact, ultra-efficient open weights like Meta’s Llama 3.2 (1B and 3B) and Alibaba’s Qwen 2.5 (3B). These models run fast on standard laptops without requiring monster GPUs, producing vivid, descriptive bedtime imagery in milliseconds.
The Sloth Prompt Engineering Framework:
- Traditional LLM prompts generate brisk, action-oriented responses. SlothSnooze uses a specialized prompt structure that forces low informational density, deliberate sentence rhythms, sensory grounding (tactile moss, nocturnal mist, slow sway), and an absence of any plot conflict.
Air-Gapped Atmospheric Fallback Engine:
- To make sure SlothSnooze works when camping off-grid or on an airplane with zero dependencies running, we implemented a procedural atmospheric generator that weaves randomized sloth folklore and sensory passages into a cohesive sleepy narrative.
Pure Procedural Audio Synthesis:
- Rather than streaming gigabytes of repetitive MP3 loops, SlothSnooze uses the browser’s Web Audio API to generate all sounds mathematically (pink noise algorithms, low-pass filter sweeps, and dual-oscillator binaural beats).

Why Does Open Innovation Matter?

Open innovation is not just an implementation detail for SlothSnooze—it is the only reason the project can exist:

Bedside Privacy & Sanctuary (Air-Gapped AI)
Your bedroom is your most vulnerable, intimate environment. Commercial sleep apps like Calm or Headspace rely on closed cloud services, and smart bedside speakers keep active microphones connected to corporate cloud servers. With open-weight models running on your local machine, you can switch your WiFi completely off. No audio leaves your machine, no sleep habits are harvested, and no advertising profiles are built.
The Death of Subscription Fatigue ($0 Forever)
Commercial sleep and meditation apps have turned basic human rest into a monthly rent-seeking scheme ($70–$100/year). Because open-source AI and open-weight models have no per-token API charges and zero recurring subscriptions, SlothSnooze is 100% free to run forever.
Independence from Cloud Outages & Latency
When you are on a red-eye flight across the Atlantic or staying in an off-grid cabin in the woods, closed APIs like OpenAI or Anthropic fail completely. Because SlothSnooze executes local weights on-device, it operates seamlessly wherever you sleep.
Total Model Freedom & Customization
With closed models, you are stuck with whatever safety filters and conversational tone the cloud provider dictates. With open weights, you can run SmolLM on a tiny handheld, Llama 3.2 on a MacBook Air, or fine-tune your own sloth bedtime narrator without asking permission from anyone.

My Agent Session

This project was built pair-programming with Antigravity, an advanced coding agent from Google DeepMind. The session involved:

Conceptualizing the International Sloth Day theme and lazy user experience.
Generating the bespoke digital artwork (sloth_snooze_hero.jpg and sloth_mascot_avatar.jpg) using text-to-image tools.
Architecting the Python FastAPI backend with local Ollama auto-discovery.
Implementing the procedural Web Audio synthesis engine with real-time binaural brainwave modulation and audio ducking.
Designing a bioluminescent dark-mode interface with an interactive starfield and breathing guide.

The Lazy Payoff

Let's measure the actual lazy return on investment:

Time saved from doomscroll insomnia: ~45 minutes per night.
Mental effort required: Exactly 1 mouse click on "Lazy Sleep Mode".
Money saved: ~$84.00/year in commercial app subscriptions avoided.
Energy Conserved: 99.4% (Barnaby the Sloth approved).

Prize Categories

Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
Best Use of Local Inference / Open-Weight Models
Best Offline / Air-Gapped Application

TrailSense: An Offline Nature Companion Built to Keep You Off Your Phone

Tatvagna Tandel — Sun, 11 Oct 2026 13:50:16 +0000

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

What I Built

Have you ever gone for a hike to clear your head, heard a bird singing high up in the trees, pulled out your phone to look it up, and immediately run into a "No Internet Connection" spinner? Or worse , you open an app, and before you can even identify what you saw, it wants you to create an account, accept location tracking, and look at a social feed?

Most outdoor tech ends up pulling you deeper into the digital world when all you wanted was to be in the physical one.

I built TrailSense because I wanted the exact opposite: an app where the screen is the shortest part of the experience.

TrailSense is a lightweight, offline first web companion for hikers, runners, and nature lovers. You pull it out, spend 3 to 5 seconds identifying a sound or leaf, and the app prompts you with a quick real-world observation before telling you to put your phone right back in your pocket:

TrailVoice (Avian Audio): You tap record, and it runs a real-time spectrogram analysis right on your device to identify what bird is singing (like an American Robin, Chickadee, or Barred Owl).
FloraLens (Foliage & Plant Vision): Snap a quick photo of a mystery leaf or wildflower, and it tells you what it is, whether it's safe (e.g., flagging Poison Ivy), and how it fits into the local ecosystem.
"Look Up & Touch Grass" Prompts: Instead of endless scrolling, every result gives you a tactile cue like looking up 20 feet into the canopy to spot wing movement, or checking leaf edges.
Sensory Outdoor Quests: Quick, screen free mini challenges (like doing a 60-second silent soundscape or feeling tree bark textures) that nudge you to pay attention to your surroundings.
Private Field Journal: Logs what you found into local browser storage with zero tracking, and lets you export a clean Markdown log whenever you're back home.

Demo

Live Site: https://touch-grass12.netlify.app

Want to test it right now without leaving your desk?
I know judges and developers might be testing this late at night or indoors, so I added:

1 Click Bird Song Samples: Click on cards like American Robin or Black-capped Chickadee to hear authentic bird call synthesis and watch the live spectrogram canvas classify it in real time.
Field Photo Samples: Click on the high res photos (Alpine Wildflower, Autumn Sugar Maple, Robin) to test the camera classifier.
Try Airplane Mode: Turn off your Wi-Fi or set Chrome DevTools Network to Offline. The app keeps running without a hiccup because everything happens right on your device.

Here’s how the pieces fit together:

index.html: Responsive, dark forest glassmorphic interface that's easy on the eyes in direct sunlight or shade.
js/ai-audio.js: Real time Web Audio API FFT decomposition and acoustic signature matching.
js/ai-vision.js: In browser canvas computer vision for leaf venation, color distribution, and plant safety checks.
js/audio-synth.js: Web Audio oscillator and filter synthesizer that mimics wild bird frequencies for testing.
js/offline-store.js: Privacy-first IndexedDB journal with Markdown export.
sw.js: Service worker caching everything so it works off grid.

How I Built It

The main design challenge was simple: no servers allowed. If this app needed a cloud backend, it would be useless on the trail.

Here’s what powers it:

Acoustic Signal Processing (Web Audio API)
When listening through the mic, TrailSense taps into the browser’s native AudioContext with an FFT size of 1024. It breaks incoming audio into frequency slices, calculates spectral centroids, and visualizes them on an HTML5 canvas. It matches frequency ranges and cadence against avian acoustic profiles (for example, identifying the distinct 2.2–3.4 kHz harmonic rhythm of an American Robin).
In-Browser Computer Vision
Instead of shipping photos to an expensive third party vision API, TrailSense draws captured frames onto an offscreen canvas. It analyzes chromatic balance, edge density, and color histograms client-side to classify native foliage and check safety guidelines.
Progressive Web App & Caching
A custom Service Worker caches all scripts, assets, and audio routines using the Cache API. Once you open the page once, it is stored locally. You could be miles into a mountain trail with zero cell bars, and it will load instantly.

Why Does Open Innovation Matter?

Building this reinforced for me why open source AI is so important:

Cell towers don't follow hiking trails: When you’re deep in a state park or camping off-grid, proprietary cloud APIs like OpenAI or Google Vision simply don't work. Open, client-side tools aren't just an alternative here they are the only way to build reliable outdoor tools.
Nature should be private: You shouldn't have to surrender your GPS coordinates, microphone recordings, and camera roll to a tech company just to learn the name of a flower. Keeping everything in the browser respects people's privacy.
Zero server costs, zero barriers: Because this runs on the user's hardware, hosting costs are literally zero. Anyone can fork it, host it on GitHub Pages or Netlify, and share it with community hiking clubs, scout troops, or schools without worrying about monthly API bills.

My Agent Session

I built and iterated on TrailSense using Antigravity AI IDE, pairing up with the agent to design the DSP audio pipeline, synthesize authentic bird acoustics for the demo, and style the dark-forest UI. It made going from concept to a deployed offline PWA surprisingly fast.

Prize Categories

Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
Edge AI & Local Inference Track
Privacy & Open Web Track