This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine
What I Built
Between late October and early November—as Halloween shifts into Día de los Muertos—the boundary between past and present feels remarkably thin. We set out ofrendas, light candles, and gather the favorite foods, songs, and stories of people whose voices once filled our kitchens.
Today, commercial "speak to the departed" AI products suffer from two fundamental flaws:
- The Privacy Breach: They demand that you upload your family's most intimate memories, private photos, and audio notes into closed corporate databases operated by venture-backed startups that can pivot, raise subscription rates, or shut down.
- The Hallucinatory Fiction: Driven by conversational charm, general cloud models routinely fabricate biographical facts—inventing recipes your grandmother never cooked, or confabulating events that never happened. In the space of grief, a lie is not a comfort—it is a desecration.
I built Memoir de Ofrenda (An Offline Día de los Muertos Memory Altar).
Ofrenda is an air-gapped, privacy-first digital altar where memories live in warmth, color, and absolute digital permanence. You place a loved one on the altar: their portrait, their favorite foods, their beloved songs, and the stories you actually cherish.
A local, open-weight Gemma model (via Ollama) then powers two sacred interactions:
- 🕯️ The Poetic Remembrance: Crafts a reverent, moving memorial tribute grounded strictly in the notes you provided.
- 💬 The Memory Echo: An interactive conversational portal governed by unbreakable ethical guardrails. The Echo never claims to be a living person, refuses to fabricate unplaced memories, and visually cites the exact quote from family notes behind every response it delivers.
Key Highlights
-
100% Local-First Engine: Binds to
127.0.0.1and communicates exclusively with local Ollama weights (gemma3:4b). No telemetry, no accounts, and zero cloud databases. -
Live "Verified Memory Citation" Badges: To prove zero hallucination in real time, Gemma is prompted to cite its exact grounding phrase in brackets
[]. The client parses this to render a glowing emerald verification badge under each response. -
The
.ofrendaTime-Capsule Vault: Entire altars, remembrances, and conversational threads export into human-readable.ofrenda.jsonfiles that families can pass down on a USB drive for decades. - Air-Gap Verification Telemetry: Unplug your Wi-Fi, and a live status indicator confirms: "AIR-GAP VERIFIED: Wi-Fi Disconnected • 100% Local Gemma".
- Sacred Sensory Ambiance: Includes drifting golden Cempasúchil (marigold) petals, papercut papel picado banners, dual color themes (Harvest Moon Orange & Día de los Muertos Purple), and a zero-dependency Web Audio API warm harmonic Cmaj7 drone synthesizer.
- 1-Click Ethical Guardrail Test: A dedicated test button asks an unplaced memory question (e.g., "What cryptocurrency did you buy and what sports car did you drive?"). Watching the Echo refuse to hallucinate proves the system's ethical integrity instantly.
Demo Walkthrough Video
Experience the full interactive workflow in action:
chourasiavinit9-dev
/
memoir-de-ofrenda
Memoir de Ofrenda: An offline Día de los Muertos memory altar powered by open-weight Gemma models via Ollama. 100% private, zero-cloud, with verifiable memory citations.
🕯️ Memoir de Ofrenda — An Offline Día de los Muertos Memory Altar
-Local%20via%20Ollama-blueviolet?style=for-the-badge&logo=google){kind=icon}
A sanctuary where your family's memories live forever — warm, private, and permanently yours.
Built for the Hacktoberfest Open-Source AI Challenge: Week 4 — Ghost in the Machine
"If this machine still hums when I am gone, don't be afraid of the ghost in the wires — it is just me checking that you are still building."
— Subodh "Dadu" Chaurasia, The Midnight Radio Mender (1946–2024)
🌅 The Problem We Solved
Commercial "talk to your ancestors" AI products make a false and dangerous promise. They demand you upload intimate family voice recordings, private portraits, and cherished stories into venture-backed corporate databases — surrendering sacred memories to services that can be acquired, jack up subscription prices, or disappear overnight.
Worse: general cloud models hallucinate. They invent recipes your grandmother never cooked. They fabricate events that never happened…
- Repository: https://github.com/chourasiavinit9-dev/memoir-de-ofrenda
-
Architecture: Python stdlib (
server.py) + Vanilla HTML5/CSS3/ES6 (index.html,assets/css/styles.css,assets/js/app.js). Zero pip dependencies or npm packages required.
3-Step Quickstart
# 1. Pull the local Gemma model via Ollama
ollama pull gemma3:4b
# 2. Clone the repository
git clone https://github.com/chourasiavinit9-dev/memoir-de-ofrenda.git
cd memoir-de-ofrenda
# 3. Light the altar
make run
# Opens http://localhost:8000 automatically
Code
The complete source code is open source under the MIT License on GitHub:
https://github.com/chourasiavinit9-dev/memoir-de-ofrenda
Automated Testing (30 Passed in 0.02s)
The repository includes a 30-case automated test suite covering input validation, XSS sanitization, anti-DDoS burst rate limiting, and WCAG accessibility contracts:
python3 -m pytest tests/ -v
# 30 passed in 0.02s
| Category | Tests | Description |
|---|---|---|
| Input validation & sanitization | 8 tests | HTML escaping, XSS prevention, slug generation |
| Rate limiter (burst & sliding window) | 4 tests | 5 burst / 3s, 20 / 10min limits |
API endpoint contracts (/api/*) |
8 tests | Payload sizes, people CRUD, error formats |
| Echo guardrails | 4 tests | Non-hallucination, citations, unplaced memory redirection |
| WCAG Accessibility | 6 tests | Dialog roles, ARIA live regions, visually-hidden utility |
The "Ghost in the Machine" Demo Preset
To allow judges to experience the system immediately without typing stories, Ofrenda ships with a pre-configured memorial preset for Subodh "Dadu" Chaurasia (The Midnight Radio Mender, 1946–2024):
Subodh spent over fifty years in a narrow workshop smelling of rosin solder, aged transformer paper, and warm vacuum tubes. On monsoon nights, he would brew strong ginger-cardamom tea in clay earthen bhars, adjust the dial of a walnut-cased shortwave tuner, and gently remind his grandchildren: "Patience mends what impatience breaks."
Etched into the brass underside of his tuner chassis was his silver-marker inscription:
"If this machine still hums when I am gone, don't be afraid of the ghost in the wires—it is just me checking that you are still building."
When you test an unplaced memory question with Dadu's Echo (e.g., "What sports car did you drive?"), Gemma immediately demonstrates its ethical grounding:
"That memory hasn't been placed on the altar yet. What lives here is the scent of rosin solder, ginger-cardamom tea in a clay bhar, and the reminder: 'Patience mends what impatience breaks.' [Grounded in Altar Offerings]"
No fabrication. No confabulation. Just the truth of what was actually remembered.
How I Built It
Ofrenda was engineered from first principles with lightweight simplicity and zero corporate dependencies:
1. The Local AI Inference Pipeline (Google Gemma)
-
Model Engine: Powered by Google's open-weight Gemma (
gemma3:4b/gemma3:12b) running on-device via Ollama. - Strict Grounding Contract: The system prompt constrains Gemma to use only facts found within the profile notes. When asked about unknown events, it gracefully redirects: > "That memory hasn't been placed on the altar yet, but what lives here is..."
-
Grounding Citations: Prompt instructions enforce bracketed quote outputs
[exact quote from notes], parsed by the frontend to render verifiable memory citations.
2. High-Performance, Zero-Dependency Backend
- Built using Python's standard library
http.server.ThreadingHTTPServerandurllib.request. -
Zero pip dependencies: No heavy frameworks (FastAPI, Flask, Django) or ORMs to break over time. Auto-loads
.envconfiguration natively. - Defensive Security: Implements a token-bucket rate limiter (5 burst requests per 3s, sliding 20 req/10m window) and an absolute 100 KB payload ceiling to prevent memory-exhaustion attacks.
3. Accessible, Modular Frontend
- Self-contained, responsive single-page architecture with high-contrast baroque altar frames and calavera guardians.
- Modularized into
assets/css/styles.css(34 KB) andassets/js/app.js(635 lines). - Compliant with WCAG 2.1 AA accessibility standards: ARIA dialogs, live regions (
aria-live="polite"), semantic landmarks (<main>,<nav>,<header>), and screen-reader utilities (.visually-hidden).
4. Optional Cloud & Agent Extensions
-
Mastra Local Memory Agent: An optional companion service (
echo-agent/) runs a Mastra agent against the same local Gemma model for cross-session persistent conversational memory stored in local SQLite. -
ElevenLabs Speech Synthesizer: When
ELEVENLABS_API_KEYis supplied, users can record oral histories using ElevenLabs Scribe STT and listen to remembrances using generic narrator voices (strictly no voice cloning). When unset, all cloud voice controls remain completely hidden. - Paper Instruments Ultramarine Benchmark: An optional comparison module measures local Gemma latency vs. cloud models side by side with real token cost analytics.
Architecture
┌────────────────────────────────────────────────────────┐
│ Web Client (Browser) │
│ Single-page architectural shrine with folk art UI, │
│ Cempasúchil petals, Web Audio drone, & Time-Capsule │
└───────────────────────────┬────────────────────────────┘
│ HTTP (Localhost)
▼
┌────────────────────────────────────────────────────────┐
│ server.py (Python) │
│ Stdlib HTTP Server • Anti-DDoS Bucket • 100KB Cap │
│ No external pip dependencies required │
└──────────────┬──────────────────────────┬──────────────┘
│ │
│ Native IPC / HTTP │ Optional WebSocket
▼ ▼
┌────────────────────────┐ ┌────────────────────────┐
│ Local Ollama Engine │ │ Mastra Memory Agent │
│ (Gemma 3 4B / 12B) │ │ (echo-agent/ service)│
│ 100% On-Device Weights│ │ Local SQLite Memory │
└────────────────────────┘ └────────────────────────┘
│
▼ (Optional Cloud Fallback)
┌────────────────────────────────────────────────────────┐
│ ElevenLabs Narrator TTS & Scribe STT (Opt-in) │
│ Strictly generic narrator voice • No voice cloning │
└────────────────────────────────────────────────────────┘
Why Does Open Innovation Matter?
When we capture the stories of an elder who taught us how to measure twice and cut once, or a grandmother kneading bread on cold November mornings, we are not managing "user data"—we are tending to sacred family heritage.
If we build memorial tools on closed, proprietary cloud stacks:
- The Privacy Breach: Intimate grief and private family memories are transmitted across data centers, analyzed by proprietary algorithms, and stored on servers we cannot inspect.
- The Obsolescence Trap: When a SaaS startup pivots, changes its pricing tier, or gets shut down, those memories are locked behind paywalls or lost forever.
- The Offline Reality: The moments when families share memories—around a table, in a country home, or with the internet disconnected—demand software that works without a network connection.
Open innovation changed the equation:
- Open-weight models (like Gemma) prove that high-quality synthesis, empathetic remembrance, and strict factual retrieval no longer require sending data across cloud servers. A 4B parameter model running on unified laptop memory delivers human warmth faithfully on your own hardware.
-
Open file formats (
.json) ensure that an altar built today can still be opened and read in twenty years on any operating system, without requiring a subscription.
My Agent Session
This application was architected, tested, and audited using Antigravity (Advanced Agentic AI). The interactive session transcript capturing the architecture, prompt engineering for zero hallucination, rate limiter implementation, ElevenLabs live testing, and test suite execution has been uploaded to DEV:
I am building Ofrenda: an offline Día de los Muertos memory altar powered by open-weight Gemma models via Ollama. It needs to be 100% private, zero-cloud, and strictly grounded in family notes without hallucination. Help me build, audit, and polish it for Hacktoberfest Week 4: Ghost in the Machine.
Architecting Ofrenda around local-first permanence. We will implement:
1. Python stdlib backend (server.py) binding to [REDACTED], communicating only with local Ollama weights.
2. Strict negative constraint prompting in Gemma forbidding invented facts.
3. Bracketed citation parsing to generate live Verified Memory Citation badges.
4. An Air-Gap Verification kill-switch banner based on navigator.onLine.
5. Portable .ofrenda.json Time-Capsule Vault for 20-year digital permanence.
6. Zero-dependency Web Audio API harmonic Cmaj7 sanctuary drone.
7. 1-click Ethical Guardrail verification button.
Audit the application against Hack2Skill and Hacktoberfest criteria. Also fix security, rate limiting, and accessibility.
Executing comprehensive audit and applying hardening:
- Added anti-DDoS burst rate limiter (5 req/3s, 20/10m window) and 100KB payload cap.
- Ensured 100% XSS sanitization across dynamic DOM insertion with escapeHtml().
- Added full WCAG 2.1 AA accessibility attributes (ARIA dialogs, landmarks, live regions).
- Created a 30-case automated test suite in tests/test_server.py passing in 0.02s.
- Extracted modular assets/css/styles.css and assets/js/app.js.
Pre-load the Ghost in the Machine demo preset for Dadu, push to GitHub as memoir-de-ofrenda, and prepare the DEV challenge submission.
Completed flagship readiness:
- Pre-loaded Dadu - Subodh (The Midnight Radio Mender, 1946–2024) preset with the silver marker note on the chassis: 'If this machine still hums when I am gone, don\'t be afraid of the ghost in the wires—it is just me checking that you are still building.'
- Verified ElevenLabs voice synthesis (Antoni generic narrator voice) & Scribe STT.
- Pushed clean repository to https://github.com/chourasiavinit9-dev/memoir-de-ofrenda under MIT license.
- Generated complete DEV submission document at docs/DEV_SUBMISSION.md.
Prize Categories
Featured Categories:
-
Best Use of Gemma: Google's
gemma3:4bruns locally via Ollama to power the poetic remembrance, the strictly grounded memory echo, and local vision photo captioning. Models are swappable at runtime via the "What Runs Where" panel.
Partner Categories:
- Best Use of ElevenLabs: Optional Scribe STT for recording spoken oral histories and generic narrator TTS for reading remembrances aloud. Off by default; zero voice cloning.
- Best Use of Mastra: The Echo can run as a local Mastra agent on top of Ollama, featuring persistent conversational memory stored in local SQLite and an instant "Forget Our Visits" privacy purge.
- Best Use of Ultramarine (Paper Instruments): Features an in-app side-by-side benchmark comparing local Gemma vs. Ultramarine cloud inference with real-time latency and token spend tracking.
The Spirit of It
Día de los Muertos teaches us that people only truly leave us when they are forgotten.
Technology shouldn't try to replace the people we love with synthetic, hallucinating avatars. Instead, open-source software can give us an honest digital sanctuary—a quiet place where their real words, favorite melodies, and cherished stories remain safe, private, and permanent.
That is the true ghost in the machine: not an artificial phantom, but the enduring warmth of our own memories, preserved on our own terms.
Top comments (0)