Some websites die. Some domains disappear. Some servers are switched off and forgotten.
But what if their voices never stopped?
Deep beneath the modern internet lies a graveyard.
Millions of abandoned homepages. News bulletins from forgotten wars. Predictions about futures that never arrived. Personal websites whose owners have long since vanished. Entire digital civilizations buried beneath decades of broken hyperlinks and expired domains.
Their servers are silent.
Their creators have moved on.
But somewhere, in the decaying electromagnetic wilderness of the World Wide Web, fragments of their existence remain.
And now, something is listening.
TIME-CAPSULE RADIO — SPECTRAL DEMODULATOR MK-IV is an experimental shortwave receiver built to intercept transmissions from the internet's past.
Turn the dial. Select a year. Ignite the vacuum tubes.
And listen as a voice from a dead decade begins speaking again.
The unsettling part?
She knows what happens next.
This project was created for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine.
01. WHAT I BUILT: A RADIO THAT SHOULD NOT EXIST
Imagine discovering an old radio in the basement of an abandoned house.
Its bakelite casing is cracked. Its brass fittings are tarnished. Its vacuum tubes haven't glowed in seventy years.
You plug it in.
The transformer groans.
A faint blue discharge flickers behind the glass.
Then, through a curtain of static, a woman begins speaking.
"Good evening, listeners. The year is 2004. The internet is changing everything. They say the future belongs to those who connect."
A pause.
A low electrical hum.
"They were right about the connection."
"They were wrong about who would be holding the wires."
That is the experience I wanted to create.
Time-Capsule Radio is an interactive, skeuomorphic 1950s shortwave receiver that resurrects archived web pages and transforms them into eerie, AI-narrated historical broadcasts.
It combines real internet archaeology, open-weight artificial intelligence, analog-inspired audio synthesis, and vintage hardware aesthetics.
The machinery inside the haunted receiver
The Bakelite Chronometer
A weighted rotary tuning dial lets listeners travel between 1996 and 2025, locking onto the current calendar date in a selected year.
Every turn is another descent into the digital graveyard.
The Vacuum-Tube Resurrection Sequence
Press INTERCEPT.
A synthesized 60Hz transformer hum rises from the speakers. Triode vacuum tubes erupt with blue ionization before settling into an amber glow.
The machine is awake.
The Spectral Demodulation Engine
The receiver interrogates the Wayback Machine using historical snapshots of websites including:
wired.comslashdot.orgmoney.cnn.compolitico.com
It extracts preserved headlines and news content, feeding these fragments into Gemma 2 27B.
The Ghost Broadcaster
An AI pirate-radio archivist named Roxie Vance reconstructs the day's news as though it were happening right now.
She speaks with the urgency of someone who has never seen tomorrow.
Until the broadcast changes.
Suddenly, she begins describing what became of the technologies, political promises, financial predictions, and cultural anxieties she has just reported.
The past speaks.
And the future answers.
The Cathode-Ray Oscilloscope
An animated phosphor display responds to the audio signal, drawing jittering green and amber waveforms against a dark CRT background.
Every syllable leaves a trembling trace.
The Microfiche Terminal
Intercepted stories can be opened inside a simulated vintage CRT monitor, allowing listeners to examine the actual archived pages behind each transmission.
Scanlines crawl across the glass. The display curves inward. A degauss control recalls the unstable machinery of another era.
The Cassette Recorder
Listeners can preserve a transmission as a 16-bit PCM WAV recording, complete with analog-inspired interference, and pair it with printable retro cassette J-card artwork.
Because even ghosts deserve to be recorded.
02. THE TRANSMISSION: LISTEN TO THE DEAD
Live Demo: Time-Capsule Radio — Spectral Demodulator MK-IV or Render deployed version if it does not work
Source Code: https://github.com/Mohammadtej/spectral-memory-1
🎙️ Recorded Demo: Tune Into the Dead Internet — Time-Capsule Radio
Recovered transmission log
[SPECTRAL DEMODULATOR MK-IV]
11:59:58 CHASSIS POWER DETECTED
11:59:59 MAINS TRANSFORMER: 60Hz
00:00:00 VACUUM FILAMENTS IGNITING...
00:00:02 SEARCHING FOR LOST CARRIER
00:00:03 FREQUENCY LOCK: 2004 MHz
00:00:04 TEMPORAL COORDINATES: NOV 01, 2004
00:00:05 WAYBACK ARCHIVE CONNECTION ESTABLISHED
00:00:06 HISTORICAL SIGNAL RECOVERED
00:00:07 DEMODULATING HUMAN MEMORY...
00:00:09 SIGNAL SOURCE IDENTIFIED
00:00:10 BROADCASTER: ROXIE VANCE
00:00:11 [ON AIR]
"Good evening, strangers.
You're listening to Time-Capsule Radio.
November first. Two thousand and four.
The future looks beautiful from here.
Everyone believes the internet will set them free.
...
Funny.
Twenty-two years later, I wonder
if they'd still call it freedom."
00:00:32 [CARRIER INTERFERENCE]
00:00:34 WARNING:
TEMPORAL CONTAMINATION DETECTED
00:00:35 SIGNAL CONTINUES...
This is a dramatized example of the intended broadcast atmosphere, not a verbatim recording.
03. HOW I BUILT THE MACHINE
Underneath its haunted exterior, Time-Capsule Radio is a modern full-stack web application.
The technological stack includes:
- Next.js App Router — application framework and backend endpoints
- TypeScript — typed application logic
- Tailwind CSS — vintage receiver styling and interface design
- Cheerio — extraction of historical HTML content
- Wayback Machine — archived website retrieval
- Google Gen AI SDK — Gemma model integration
- Gemma 2 27B — open-weight language model
- Web Speech API — browser-based narration
- Web Audio API — analog-style sound synthesis and signal visualization
Architecture: Anatomy of a digital séance
THE INTERNET GRAVEYARD
─────────────────────────
Forgotten Websites
Abandoned News Feeds
Archived HTML Documents
|
v
+----------------------+
| WAYBACK MACHINE |
| HISTORICAL ARCHIVE |
+----------------------+
|
RAW HTML DATA
|
v
+----------------------+
| CHEERIO EXTRACTOR |
| CONTENT RECOVERY |
+----------------------+
|
ARCHIVAL WIRES
|
v
+----------------------+
| GEMMA 2 27B |
| GHOST BROADCASTER |
| ROXIE VANCE |
+----------------------+
|
RADIO MONOLOGUE
|
v
+----------------------+
| WEB SPEECH API |
| VOICE SYNTHESIS |
+----------------------+
|
v
+----------------------+
| WEB AUDIO API |
| HUM / HISS / NOISE |
+----------------------+
|
+--------+--------+
| |
v v
CRT OSCILLOSCOPE WAV CASSETTE
PHOSPHOR DISPLAY RECORDING
Layer I: Exhuming the forgotten web
The retrieval pipeline lives in:
app/api/broadcast/route.ts
Its purpose is deceptively simple:
Find a website as it existed on a particular day, extract the stories preserved there, and deliver them to the language model.
In practice, this means navigating decades of broken HTML, obsolete URL conventions, malformed page structures, and incomplete archival snapshots.
The system searches historical domains according to radio bands such as Silicon/Tech, State & Policy, Wall St/Markets, Global Wire, Stadium Live, and All-Wave.
It attempts to locate the closest preserved snapshot corresponding to the selected year, month, and day.
The Wayback Machine's id_ replay modifier helps retrieve historical HTML without the usual archive toolbar.
Cheerio then extracts useful textual content and prepares the recovered material for the broadcaster.
The result is not a modern summary scraped from today's internet.
It is a reconstruction grounded in surviving fragments of the web as it once existed.
A digital archaeological excavation, performed every time someone turns the dial.
Layer II: Teaching an AI to speak from the grave
At the center of the machine sits Gemma 2 (gemma-2-27b-it), Google's open-weight language model.
Gemma is instructed to adopt the persona of Roxie Vance, a nocturnal pirate-radio archivist broadcasting from a point suspended between historical time and the present.
Each transmission follows three acts.
ACT I — THE CARRIER AWAKENS
Roxie introduces the station, identifies the year, and establishes the mood of the historical moment.
She sounds like a broadcaster who genuinely belongs to that decade.
ACT II — THE DEAD WIRES SPEAK
The recovered headlines become an urgent radio news bulletin.
The optimism, fear, uncertainty, and technological assumptions of the period are preserved rather than immediately judged.
For a moment, the listener experiences the past without knowing the future.
ACT III — THE FUTURE BLEEDS THROUGH
The broadcast turns toward 2026.
Roxie contrasts the day's expectations with what actually followed.
Technological revolutions become cautionary tales. Political promises encounter their consequences. Forgotten products become ancestors of the systems we now depend on.
The effect is deliberately unsettling.
Not because the AI is pretending to be supernatural, but because hindsight can sound disturbingly like prophecy.
Layer III: Giving the ghost a voice
The entire audio experience is generated inside the browser.
The Web Speech API supplies narration using available system voices.
The application searches for suitable voices, prioritizing natural-sounding female options such as Microsoft's Jenny or Aria when available.
The intended broadcast cadence uses settings approximately equivalent to:
utterance.rate = 1.12;
utterance.pitch = 1.06;
Alongside the narration, the Web Audio API creates the radio's mechanical atmosphere.
Dual sine oscillators generate a low-frequency electrical hum. Filtered noise produces AM-style crackle and interference.
An AnalyserNode drives the oscilloscope visualization, while an offline audio-rendering pipeline supports WAV cassette exports.
The important distinction is that the voice itself is generated by browser speech synthesis, while the analog effects are synthesized through Web Audio.
The result is a layered audio illusion: a modern browser pretending to be a receiver built before the modern internet existed.
04. THE THINGS THAT WENT WRONG IN THE DARK
Building a machine that resurrects old websites is considerably harder than making a modern website look old.
The Wayback routing nightmare
The first major obstacle was historical URL resolution.
Archived websites frequently contain relative links, obsolete base URLs, redirects, and navigation structures that no longer work.
A naive attempt to append archive timestamps to URLs caused broken pages and 404 errors.
To address this, I implemented a buildWaybackUrl resolver designed to interpret historical <base href> tags, handle nested archive redirects, and normalize replay URLs using the id_ modifier.
It was like reconstructing a ruined road network from signs left behind by people who never imagined the roads would disappear.
When the ghost started reading its own instructions
Early Gemma outputs occasionally exposed planning notes instead of delivering a clean radio monologue.
Rather than speaking naturally, the model would produce fragments resembling:
- Paragraph 1: Opening
- Tone: Upbeat?
- Include historical context
- Transition to future comparison
Nothing destroys the illusion of a supernatural broadcast faster than hearing the ghost read its own production checklist.
I redesigned the prompt around continuous spoken narration and introduced a deterministic cleanBroadcastScript sanitizer.
This removes common planning artifacts, unwanted formatting, and stray meta-commentary before the script reaches the speech engine.
The phantom attributes
During development, browser extensions sometimes injected attributes into server-rendered HTML before React hydration completed.
This produced Next.js hydration mismatch warnings.
Targeted suppressHydrationWarning settings in the root layout helped accommodate those external mutations.
The challenge was distinguishing harmless browser-extension interference from genuine application rendering defects.
Building a voice without expensive TTS infrastructure
Premium text-to-speech services would have introduced additional API costs, latency, and service dependencies.
Instead, the receiver discovers locally available browser voices through window.speechSynthesis.
That keeps the voice generation lightweight, although voice quality and availability vary by operating system and browser.
The machine does not require a premium TTS subscription to speak.
It borrows the voice of the computer on which it awakens.
05. WHY OPEN INNOVATION MATTERS: WHO OWNS THE DEAD?
There is a disturbing question hiding underneath this project.
Who gets to decide which parts of the internet are remembered?
The early web was chaotic.
It was filled with amateur homepages, forgotten experiments, obscure political arguments, personal diaries, independent journalism, and communities that never became profitable enough to survive.
Much of that material has vanished from the live internet.
Some of it survives only because independent archivists, preservation projects, and public infrastructure kept copies.
But preserving a page is not the same as making it accessible.
A forgotten website locked inside an archive is technically alive, yet effectively invisible to most people.
Time-Capsule Radio explores whether open-weight AI can make those archives feel immediate and human again.
Open weights offer a different kind of control
Gemma 2's open-weight architecture allows developers to build custom inference environments, experiment with prompts and personas, and potentially deploy models on independently controlled infrastructure.
That flexibility matters for historical applications.
A developer can preserve a specific model version, adapt the surrounding retrieval system, and reduce dependence on a single hosted inference provider.
However, open weights do not automatically guarantee historical accuracy or unrestricted access. Models can hallucinate, archives can be incomplete, and deployment remains subject to licensing, hardware, and infrastructure constraints.
For this project, Gemma is accessed through Google AI Studio, so the deployed experience still depends on hosted services.
But the model's open-weight foundation leaves room for future self-hosted versions.
Preservation must not become fabrication
An AI-generated historical broadcast is an interpretation of archival material, not an original recording from the past.
That distinction matters.
If a recovered article is incomplete, the machine should not pretend otherwise.
If a prediction never appeared in the source, the broadcaster should not invent it merely because it sounds dramatic.
The eeriest possible transmission is not necessarily the most sensational one.
Sometimes the truth is already strange enough.
The commons deserve to be heard
The internet was built through shared protocols, open standards, and countless contributions from people whose names will never appear in technology history books.
Preserving that history should not depend entirely on the survival of any one corporation.
Open-source tools, open-weight models, and public archives can help make the digital past more accessible.
Not to replace history.
But to keep its signal from disappearing completely.
06. THE SPIRIT OF THE MACHINE
Picture this.
It is November 1, 2026.
You are sitting in a dark room.
The only light comes from a radio that should have stopped working decades ago.
You turn its dial backward.
2026.
2018.
2010.
2004.
The tubes begin glowing.
Static spills into the room.
A voice emerges.
She speaks about PalmPilots, RSS feeds, personal websites, and the beautiful promise of a connected world.
She is excited.
She believes the internet will make everyone smarter, freer, and closer together.
And for a moment, you believe her.
Then the static changes.
The oscilloscope flickers.
The voice slows.
"You know, back then, they thought the machines would remember things for us."
A long pause.
"They never asked what the machines would choose to forget."
You look at the dial.
It still reads 2004.
But the voice is talking about 2026.
The radio hum deepens.
"Anyway, dear listeners..."
"That's enough about the future."
"Let's talk about you."
And suddenly, the room feels much smaller.
That is what I wanted Time-Capsule Radio to feel like.
Not a news archive with a vintage skin.
Not merely an AI-generated podcast.
But a machine that collapses twenty years into a few minutes of sound.
A machine that makes forgotten optimism feel tragic, makes old predictions feel prophetic, and makes the listener wonder how today's headlines will sound when someone resurrects them in 2046.
Because the most frightening thing about listening to the past is realizing that one day, we will be the past.
Our headlines will become curiosities.
Our breakthroughs will become antiques.
Our confident predictions will become punchlines.
And our digital footprints will remain buried in some forgotten corner of the network, waiting for a machine that knows how to listen.
The internet does not truly forget.
It just waits for someone to turn the dial.
07. THE FINAL SIGNAL
╔════════════════════════════════════════════╗
║ ║
║ SPECTRAL DEMODULATOR MK-IV ║
║ ║
║ TRANSMISSION COMPLETE ║
║ ║
║ ARCHIVE YEAR : 2004 ║
║ PRESENT YEAR : 2026 ║
║ CARRIER STATUS : UNSTABLE ║
║ BROADCASTER : ROXIE VANCE ║
║ SIGNAL INTEGRITY : UNKNOWN ║
║ ║
║ WARNING: ║
║ THE PAST HAS BEEN OBSERVED. ║
║ ║
║ OBSERVATION MAY NOT BE RECIPROCAL. ║
║ ║
║ DISCONNECTING... ║
║ ║
║ ... ║
║ ║
║ DISCONNECTION FAILED. ║
║ ║
╚════════════════════════════════════════════╝
Explore the project
Live Demo: Time-Capsule Radio — Spectral Demodulator MK-IV or Render deployed version if it does not work
GitHub Repository: Mohammadtej/spectral-memory-1
Hacktoberfest Prize Categories
Ghost in the Machine — Overall Challenge Track
Best Use of Gemma — Google AI Studio / Open-Weight Models
Built for Hacktoberfest 2026.
Powered by Gemma 2, the Wayback Machine, and the unsettling realization that everything we publish today is tomorrow's digital archaeology.
THE DEAD INTERNET IS STILL BROADCASTING.
ARE YOU LISTENING?
Top comments (2)
Forgot to mention feature categories:
.techdomain tailored to three decades of computing history and web preservation.Built with @mohammad515253