This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 1 — Touch Grass.
PALABRE: What If AI's Job Was to Get You Off Your Phone?
We are getting very good at building AI that keeps us on screens.
More answers.
More recommendations.
More generated content.
More reasons to keep scrolling.
So when I saw this week's Touch Grass challenge, I wanted to try something deliberately backwards:
What if AI's job was not to keep you online, but to send you outside?
That question became PALABRE.
Your phone knows the internet. Your neighborhood knows the stories.
PALABRE is a real-world cultural treasure hunt powered by AI.
It gives you a mission, sends you into your neighborhood to talk to a real person, lets you record what you hear, analyzes that testimony, preserves the resulting knowledge, and generates the next mission from the story you just collected.
The screen is only the beginning.
The real experience happens outside it.
What I Built
PALABRE is a real-world cultural treasure hunt powered by AI.
Instead of asking the user to spend more time talking to an AI, PALABRE gives them a mission that must be completed in the real world.
The loop is simple:
MISSION
↓
GO OUTSIDE
↓
TALK TO A REAL PERSON
↓
RECORD THE STORY
↓
AI ANALYSIS
↓
PRESERVE THE KNOWLEDGE
↓
NEXT MISSION
↓
GO OUTSIDE AGAIN
The project starts with the neighborhoods of Yaoundé, Cameroon, where much knowledge is still passed through conversation rather than databases, websites, or formal archives.
A proverb.
A craft lesson.
A neighborhood memory.
A piece of advice from an elder.
A story about work, family, migration, food, language or everyday life.
PALABRE is designed to turn those conversations into a growing living oral archive.
And there is one important twist:
the next mission is generated from the story that was just collected.
So the experience doesn't simply repeat the same AI prompt.
It evolves.
The Problem I Wanted to Solve
A lot of our digital lives revolve around consuming information that already exists online.
But some of the most valuable knowledge around us is not online at all.
It lives in conversations.
A proverb repeated by an elder.
A lesson learned through years of work.
A neighborhood memory.
A craft technique.
A family story.
An expression that only makes sense when someone explains where it came from.
That knowledge can disappear without ever becoming searchable.
And there is an interesting contradiction here:
AI can generate enormous amounts of new information, while millions of real human stories remain uncaptured.
PALABRE explores a different use of AI:
not generating another story, but helping preserve a story that already exists.
The PALABRE Loop
The entire product is built around one physical loop:
MISSION
↓
GO OUTSIDE
↓
TALK TO A REAL PERSON
↓
RECORD THE STORY
↓
AI ANALYSIS
↓
CULTURAL MEMORY
↓
NEXT MISSION
↓
GO OUTSIDE AGAIN
The important part is the last step.
PALABRE does not simply analyze a conversation and end the experience.
The next mission is generated from the testimony that was just collected.
That means the quest can evolve through actual human encounters.
A conversation creates a new question.
A new question creates another conversation.
The archive grows through people, not through endless AI chat.
What the Explorer Actually Does
A PALABRE mission might ask the explorer to return to their neighborhood and ask a merchant, neighbor, craftsperson or elder a meaningful question.
For example:
What is a piece of advice you were given by your elders that still guides your work today?
The application then gets out of the way.
You leave the screen.
You find someone.
You ask.
You listen.
You record.
You come back.
That is intentional product design.
For PALABRE, a successful session should end with less screen time, not more.
Building a Real Voice-Recording Experience
The audio layer is implemented directly in the browser using the Web Audio API and MediaRecorder.
I wanted recording to feel like a proper part of the product rather than a file-upload form.
PALABRE handles:
- microphone permissions
- supported recording formats
- real-time audio-level feedback
- recording duration limits
- invalid or empty recordings
- file-size validation
- re-recording
- server-side upload validation
The current maximum recording duration is three minutes, with a 25 MB upload limit.
The flow is:
Browser microphone
↓
Audio recording
↓
FormData upload
↓
Express API
↓
Speech-to-text / AI
AI That Listens Instead of Talking
This is where PALABRE becomes more than a recording app.
The AI pipeline separates speech recognition from cultural analysis.
Conceptually:
AUDIO
↓
SPEECH-TO-TEXT
↓
TRANSCRIPT
↓
AI ORCHESTRATOR
↓
STRUCTURED CULTURAL ANALYSIS
Instead of asking an LLM for an arbitrary paragraph, PALABRE expects a structured result:
{
"transcript": "...",
"theme": "...",
"extracted": "...",
"summary": "...",
"culturalSignificance": "...",
"nextMission": "..."
}
That output becomes the bridge between the human conversation and the next physical action.
I Was Careful About Cultural Hallucinations
This part mattered more than I expected.
If an AI is analyzing a cultural testimony, inventing a plausible-sounding proverb is not a harmless mistake.
It becomes misinformation about someone's culture.
So PALABRE's analysis instructions explicitly tell the model to:
- stay grounded in the testimony
- preserve what was actually said
- avoid inventing names, dates or traditions
- distinguish direct information from inference
- avoid unsupported anthropological claims
- mark unclear speech instead of pretending to understand it
The objective is not to make the result sound impressive.
The objective is to make it faithful.
That distinction matters when dealing with oral history.
From One Human Voice to a Preserved Memory
The result turns the recording into a structured cultural artifact.
PALABRE can surface:
What was shared
The core content of the testimony.
The AI heard
The main insight extracted from the story.
Why it matters
A cautious explanation of its significance for community memory.
Preserved voice recording
The original recording can remain attached to the story.
That last part is important.
An oral archive containing only transcripts loses something fundamental.
You lose the voice.
The rhythm.
The pauses.
The personality.
PALABRE therefore treats the original recording as part of the memory itself.
The Next Mission Is Generated From the Story
This is the part I care about most.
Imagine that someone tells you how an older relative taught them a craft.
PALABRE can use that testimony to create a new mission that asks you to find another person related to that theme and ask another meaningful question.
The product therefore forms a chain:
HUMAN STORY
↓
AI UNDERSTANDING
↓
CULTURAL MEMORY
↓
NEXT PHYSICAL QUESTION
↓
ANOTHER HUMAN STORY
The AI is not the destination.
It is the mechanism that keeps the human experience moving.
The next mission comes from the story you just heard.
That's what makes PALABRE different from simply putting an LLM behind a form.
Field Proof: Did the Digital Mission Actually Lead Outside?
I wanted the project to take the challenge literally.
So PALABRE includes Field Proof.
At the beginning of a mission, the app records the starting time and can request a best-effort location.
At submission, it can calculate an approximate straight-line distance using the Haversine formula.
Only derived values are sent to the server:
time outside
distance
Raw coordinates are not stored.
And the mechanism is deliberately non-blocking.
No location?
The mission still works.
No GPS fix?
The mission still works.
The feature is there to acknowledge the physical world, not to turn PALABRE into a tracking application.
Under the Hood: An AI Orchestrator Instead of a Single Vendor
I also did not want the application permanently tied to one AI provider.
So the server contains a provider abstraction and an AI Orchestrator.
The current architecture supports:
Ollama
Qwen
DeepSeek
GLM
Gemini
The first four provide paths toward open-weight or local inference, while Gemini can remain available as a fallback depending on deployment configuration.
The application can inspect provider configuration and health, route requests according to capabilities, and fall back between configured providers.
The high-level architecture is:
┌── Ollama
├── Qwen
VOICE → STT → TRANSCRIPT → AI ORCHESTRATOR ──┼── DeepSeek
├── GLM
└── Gemini
↓
STRUCTURED ANALYSIS
↓
CULTURAL MEMORY
↓
NEXT MISSION
This means the rest of PALABRE does not need to know which model is doing the reasoning.
The intelligence layer can evolve independently.
Why Open Innovation Matters Here
This challenge could have been solved with a simple:
Browser → one closed AI API → response
That would work.
But it would also make the intelligence layer much harder to replace.
PALABRE instead keeps the AI layer modular.
That creates room to:
- swap models
- test different providers
- run local inference
- reduce vendor lock-in
- choose different inference strategies
- move more processing closer to the user over time
This matters especially for a project starting in Cameroon.
A product designed for real communities cannot assume unlimited bandwidth, expensive hardware, permanent connectivity, or identical access to cloud services.
The longer-term architecture I want for PALABRE looks like:
AUDIO
↓
LOCAL SPEECH-TO-TEXT
↓
LOCAL OPEN-WEIGHT MODEL
↓
CULTURAL ANALYSIS
↓
ARCHIVE
That is a direction rather than a claim that every deployed component is already fully local today.
The important thing is that the architecture does not make that future impossible.
The Technology Behind the Experience
The user sees a relatively simple product.
Underneath, there is quite a bit going on.
Frontend
- React 19
- TypeScript
- Vite
- Tailwind CSS
- Web Audio API
- MediaRecorder
- browser geolocation
Backend
- Node.js
- Express
- multipart audio handling with Multer
- rate limiting
- server-side AI orchestration
- structured output validation
Data
- persistent story metadata
- optional Firestore synchronization
- original voice recordings when available
- public collective oral archive
AI Layer
- provider abstraction
- capability detection
- provider health checks
- automatic fallback
- speech-to-text abstraction
- strict JSON validation
The result is a fairly simple user experience sitting on top of an intentionally decoupled architecture.
Testing the Failure Paths
A demo that works once is not enough.
I added tests around:
- provider configuration
- AI routing
- fallback behavior
- JSON sanitization
- provider health
- capability detection
- speech-to-text routing
- health monitoring
- debug tooling
The repository currently reports:
48 tests passing.
The goal was to make failure predictable rather than invisible.
For example, if a configured provider becomes unavailable, the orchestrator can move to the next available provider instead of forcing the entire explorer experience to stop immediately.
Design Was Part of the Experiment
I deliberately avoided the usual "AI startup" visual language.
No:
- glassmorphism
- purple-blue gradients
- floating AI brains
- giant dashboards
- excessive cards
- futuristic holograms
PALABRE uses a much more editorial direction:
#090909 → background
#141414 → surfaces
#242220 → borders
#FF6B35 → signal / accent
#F2EEE8 → primary text
#77736D → muted text
The typography combines:
- Syne for display headlines
- Fraunces for editorial and oral-story moments
- Plus Jakarta Sans for interface text
- a monospace style for technical metadata
The design is intentionally restrained.
The irony is important:
an application about leaving the screen should not try to maximize your attention on the screen.
Built in Yaoundé, Designed to Travel
PALABRE starts in Yaoundé, Cameroon.
That starting point matters to me because oral knowledge is not some abstract historical concept.
It exists in ordinary conversations.
In neighborhoods.
At workshops.
In markets.
In families.
On the street.
The initial use case is local, but the model is broader.
The same interaction could preserve community memory in another city, another country, or another language.
The system is designed around a universal action:
go meet someone and listen.
What I Learned
The biggest lesson was that the hard part of AI product development is often not the model call.
It is everything surrounding the model.
How do you capture audio reliably?
What happens when the microphone fails?
What happens when the model returns malformed JSON?
What happens when an AI provider is unavailable?
How do you keep the AI grounded in what a person actually said?
How do you prevent a cultural archive from turning hallucinations into "facts"?
And perhaps the most important question:
What should the human do after the AI finishes?
For PALABRE, the answer is intentionally not:
"Keep chatting."
It is:
"Go find someone else to talk to."
What's Next?
PALABRE is an MVP, not a finished cultural infrastructure.
The next iterations I want to explore are:
More local inference
A complete local voice → open-weight model pipeline for environments with limited connectivity.
Better multilingual support
Especially for the linguistic reality of Cameroon, including French, English, Pidgin, Camfranglais and eventually more local languages where reliable speech data and models permit it.
A stronger oral archive
Better browsing and discovery of collected stories while preserving the context in which they were recorded.
Better field interactions
More meaningful ways to connect missions to real locations and communities without turning the product into a surveillance system.
The larger idea is simple:
use open AI to strengthen human memory instead of replacing human interaction.
Why PALABRE Fits "Touch Grass"
The challenge asks us to build with open-source AI/open-weight AI in a way that gets people off the screen and into the world.
PALABRE takes that literally.
The best version of the product is the one where:
SCREEN TIME
↓
MISSION
↓
SCREEN OFF
↓
REAL CONVERSATION
↓
VOICE
↓
AI
↓
MEMORY
↓
NEXT MISSION
↓
SCREEN OFF AGAIN
The technology is important.
But the technology is not the point.
The point is the person standing outside the application.
Try PALABRE
🌍 Live demo: palabre.onrender.com
💻 GitHub repository: github.com/kiddsaut07/palabre
PALABRE is open source under the Apache-2.0 license.
Try the mission.
Then do the part the application cannot do for you:
put your phone down.
Go talk to someone.
Listen.
Come back.
Bring a story with you.
Final Thought
We are entering an era where AI can generate almost anything on demand.
Maybe that makes the things AI cannot generate for us even more valuable.
A person's memory.
A voice.
A neighborhood story.
A lesson passed from one human being to another.
PALABRE is my small experiment in using AI not to replace that interaction, but to send us back toward it.
Your phone knows the internet.
Your neighborhood knows the stories.
Links
🌍 Live Demo: https://palabre.onrender.com
💻 GitHub: https://github.com/kiddsaut07/palabre
AI Assistance Disclosure
I used AI tools during development and while drafting and refining this article. I reviewed the final version and kept the technical claims grounded in the actual PALABRE repository and implementation.
devchallenge #hf26challenge #Hacktoberfest #OpenSource #AI #TouchGrass #Render
Top comments (0)