build-small-hackathon/roadb-other-screen
Road B: The Other Screen
Talk to a fictional version of yourself who chose differently.
Road B is a small-model interactive fiction experience for the Build Small Hackathon, Chapter Two: An Adventure in Thousand Token Wood.
You name a fork in your life. The app opens a cinematic Other Screen and lets you speak with a fictional alternate self: the version of you who took Road B.
It is not prediction, therapy, or advice. It is a strange mirror.
Submission links
Hugging Face Space: https://huggingface.co/spaces/build-small-hackathon/roadb-other-screen
Demo video: https://kilickursat.github.io/roadb-modal/
Social post: https://x.com/kursatkilic_/status/2065290557921091904?s=46
What the app does
Road B turns a life fork into a limited-signal ritual.
The user can:
- describe a decision point
- invoke a fictional Road B self
- chat with that alternate self
- collect Echo Artifacts
- unlock a Final Transmission
- open the Paradox Box
- download a souvenir card from the road not taken
- export a synthetic-style trace of the interaction
The core loop is:
Name the fork
→ Tune the Other Screen
→ Meet Road B
→ Collect Echo Artifacts
→ Unlock Final Transmission
→ Open the Paradox Box
→ Save the Souvenir CardEcho Artifacts
Road B is not just a chat app. After the first transmission, the user can collect artifacts from the alternate life:
- Cost Ledger
- Beauty Ledger
- A Typical Tuesday
- The Unsent Letter
- The Moment It Split
After three Echo Artifacts, the Final Transmission unlocks.
Paradox Box
The Paradox Box is the final souvenir ritual.
It turns the Road B session into a visible artifact built from the actual generated content:
- universe ID
- Road B self
- signal state
- gift
- cost
- collected Echo Artifacts
- final transmission
- strongest Road B line
The card is not random. It belongs to the specific Road B session the user just created.
Runtime
Road B is hosted as a Hugging Face Gradio Space.
The Hugging Face Space handles:
- custom cinematic UI
- Gradio app shell
- Road B session state
- Echo Artifacts game loop
- Paradox Box UI
- souvenir card rendering
- trace export
- public Space hosting
Model inference runs on Modal GPU using:
unsloth/Qwen3.5-9B-GGUFQwen3.5-9B-Q4_K_M.ggufllama-cpp-python- llama.cpp runtime
The Hugging Face Space calls our Modal endpoint for model inference. The Modal endpoint runs the GGUF model through llama.cpp.
Road B does not call hosted foundation-model APIs such as OpenAI, Anthropic, Gemini, or Hugging Face Inference Providers. Modal is used as GPU compute for our own llama.cpp endpoint.
Model
Model repo: unsloth/Qwen3.5-9B-GGUF
Model file: Qwen3.5-9B-Q4_K_M.gguf
Runtime: Modal GPU + llama.cpp via llama-cpp-python
Mock mode: falseThe model is 9B parameters, within the hackathon’s small-model limit.
Why Road B fits the judging criteria
Genuinely delightful
Road B feels like a small strange machine: portal hero, signal chamber, alternate-self chat, Echo Artifacts, Final Transmission, Paradox Box, and a downloadable souvenir card.
AI is load-bearing
Without the model, there is no alternate self, no transmission, no artifact, no final message, and no meaningful souvenir card.
The AI is the experience.
Originality of concept
Road B is not a generic chatbot. It is a fictional machine for talking to the life beside yours.
Polish of the Gradio app
The app uses a custom cinematic frontend, active navigation, Echo Artifacts, a visible Paradox Box, trace export, and Modal-powered Qwen inference for smoother judging.
Bonus badge proof
Off-Brand
Road B uses a custom cinematic frontend instead of the default Gradio look.
It includes:
- portal hero
- animated signal atmosphere
- alternate-self chat chamber
- Echo Artifacts
- Final Transmission unlock
- Paradox Box
- downloadable souvenir card
- active navigation and menu actions
Llama Champion
The model runs through llama.cpp using llama-cpp-python.
The llama.cpp runtime runs on Modal GPU.
Sharing is Caring
A synthetic public trace is included here:
samples/public_trace_sample.json
The live app also supports trace export.
Field Notes
A short build report is included here:
docs/FIELD_NOTES.md
Off the Grid note
Road B does not claim the Off the Grid bonus in the final Modal version.
The app uses an open GGUF model through llama.cpp, but inference runs on Modal GPU compute rather than fully inside the Hugging Face Space.
Files
app.py
index.html
README.md
requirements.txt
assets/favicon.svg
assets/hero-reference.png
docs/FIELD_NOTES.md
samples/public_trace_sample.json
thumbnail.pngEnvironment variables
The Hugging Face Space expects these secrets:
MODAL_QWEN_URL
MODAL_QWEN_TOKEN
MODAL_TIMEOUTRecommended values:
MODAL_TIMEOUT=900
MAX_TOKENS=850
MODEL_FILENAME=Qwen3.5-9B-Q4_K_M.ggufHealth check
The running app exposes:
/healthA healthy Modal configuration should show:
{
"runtime": "Modal GPU + llama.cpp",
"modal_qwen_enabled": true,
"modal_qwen_url_set": true,
"mock_mode": false,
"max_tokens": 850
}Demo
Demo video page:
https://kilickursat.github.io/roadb-modal/
The demo shows:
Hero
→ sample fork
→ Qwen-generated Road B self
→ Echo Artifacts
→ Final Transmission
→ Paradox Box
→ Souvenir CardSafety and privacy note
Road B is speculative interactive fiction. It is not a medical, legal, financial, psychological, or crisis-support tool.
User text is sent from the browser to the Hugging Face Space backend, then to the project’s Modal inference endpoint so the Qwen GGUF model can generate a response.
The public trace in samples/public_trace_sample.json is synthetic and does not contain real user data.
