A newer version of the Gradio SDK is available: 6.20.0
Demo video kit β shot list & script (target: 100β120 seconds)
Field-guide intel: "Storytelling counts as much as the build β no humility, sell it." Video may be uploaded directly to the Space as a file. Record at 1080p+; phone-filming-a-screen is fine for the human shots.
Ready-made assets
| Asset | Where |
|---|---|
| Live Space | https://build-small-hackathon-mystery-mail-guardian.hf.space |
| Language deep-links | append ?lang=hi / ?lang=es / ?lang=ja |
| Local mock with auto-demo | $env:GUARDIAN_MOCK="1"; .venv\Scripts\python.exe app.py then /?lang=ja&autorun=1 (b-roll; never burns GPU) |
| 4-language speech WAVs | modal_artifacts/speech_{en,hi,es,ja}.wav (48 kHz VoxCPM2) |
| UI stills (b-roll) | docs/ui/*.png |
| Synthetic letters to print | run .venv\Scripts\python.exe scripts\make_samples.py, print assets/samples/{bill,scam}.png |
Shot list (the five beats the judges were told to look for)
1. The person (0:00β0:20) β beat (a) Real person, real kitchen table, real mail in hand. One line of setup, in their own words: what confusing or scary mail looks like to them.
"This is for [name]. Letters like these are how people like her get robbed."
2. The scam catches fire (0:20β0:50) β beat (b) Phone camera on the printed gift-card scam letter β "Read my letter" β the envelope-flap loader β three paper sheets land with the red warning stamp. Read two of the reasons out loud. Then tap π β let VoxCPM2 speak a sentence.
"Twenty seconds, fully offline: warning stamp, the exact tricks named, and the scammer's phone number is nowhere on this screen."
3. The safety layer story (0:50β1:10) β beat (c) One sentence over B-roll (BUILD_LOG on screen or the warning card):
"In live testing the model once suggested contacting the tax bureau and paying soon. Our safety layer is code, not a prompt β it threw that advice away, kept the warning, and always appends: verify through a number YOU already trust."
4. Built for elders, down to the font (1:10β1:30) β beats (d) + easter egg
Quick cuts: 56px language buttons (English | ΰ€Ήΰ€Ώΰ€¨ΰ₯ΰ€¦ΰ₯ | EspaΓ±ol | ζ₯ζ¬θͺ) β
tap ΰ€Ήΰ€Ώΰ€¨ΰ₯ΰ€¦ΰ₯, whole UI flips β tap ζ₯ζ¬θͺ and the Gen X Soft Club easter egg
melts the screen into pastel Y2K (still WCAG-AA, say it!).
"The body font is Atkinson Hyperlegible β designed by the Braille Institute for low-vision readers. An elder-readability font in an elder-readability app. And yes β Japanese has an easter egg."
5. The close (1:30β1:50) β beat (e) Footer on screen (model colophon visible):
"Every model is OpenBMB β 3.3 billion parameters total, smaller than the cap by ten times, 100% local. The letter never leaves the device. All our GPU testing ran on Modal β including the eval that caught a privacy bug before any human did. Mystery-Mail Guardian: small models, watched closely, guarding the people we love."
Recording gotchas
- ZeroGPU cold start: do one throwaway analysis before recording shot 2 so the real take runs warm (~20β40s).
- The easter egg toggles on selecting ζ₯ζ¬θͺ β switch back to English first so the flip is visible on camera.
- Don't show the
?autorunURL on screen (mock-only tooling). - The disclaimer line must be visible in at least one shot (it always is β footer). Keep it: judges read it as a feature, because it is one.