"You and I — defying gravity."
Track: Digital Equity & Inclusivity Submission: SUBMISSION.md · Pitch deck · 3-min video Reference: Architecture · Rules compliance Live demos: Judges quickstart · Colab notebook · Web UI Website video: CommunityOne Submission Web MP4
When I asked Google for a list of every city, county, and school-district website in my state, the answer was "technically impossible." The federal government charges $500,000+ for "public" health data. Nonprofit aggregators paywall basic registries behind thousand-dollar subscriptions. Where data does exist, it's buried in expert jargon — frame analysis, fiscal incidence — a moat designed to keep the 99% out of the conversation.
CommunityOne is now the only place where you can find a free, complete list of local-government websites for every state in the nation — and the only truly free public platform that fuses meeting minutes, donor dollars, and legislation to measure what's actually working in local democracy.
Built on Gemma 4, it speaks Southern "fixin' to" and Northern "about to." It reads a 1990s photocopied zoning amendment and a scholar's frame analysis with equal fluency, then maps both into one view anyone can understand. No PhD required. No toll.
Size shouldn't dictate importance. Information shouldn't be a luxury good. CommunityOne is the equalizer.
Point CommunityOne at a folder of agendas, minutes PDFs, audio recordings, and contact photos for any jurisdiction. Gemma 4 returns:
| Output | Who it serves |
|---|---|
| Plain-text OCR of scanned PDFs (Demo 1) | Residents decoding 1990s photocopies of zoning amendments |
| Per-page token-budget routing (Demo 2) | Cost control — tables get 1,120 image tokens, body text gets 64 |
| Policy deconstruction with visible reasoning (Demo 3) | Journalists & researchers — who steered each decision, and why |
| Audio chunking + policy-drift detection (Demo 4) | Citizens — what changed between the agenda and the actual vote |
| Plain transcripts in any language (Demo 4a) | Non-English-speaking residents |
| Contact-photo enrichment (Demo 5) | Representation audits |
| Cross-jurisdiction embedding clusters (Demo 6) | Researchers — is Tuscaloosa passing the same ordinance as Big Timber? |
| ShieldGemma safety review (final step) | Every output checked for hallucination & stereotyping before publication |
The pipeline runs against two real reference jurisdictions already ingested:
- Tuscaloosa County, Alabama — mid-size urban county
- Big Timber & Sweet Grass County, Montana — rural, population ~3,500
Geography is encoded in the folder layout (<STATE>/<scope>/<jurisdiction>/), so FIPS codes, county names, and postal codes fall out of the path. Point at any other jurisdiction and the same code runs.
- Native multimodality — PDFs and audio go in as bytes. No separate OCR. No separate ASR.
- Adjustable media-resolution token budget — HIGH (~1,120 tokens) for ledgers, LOW (~64) for body text. The model itself exposes cost-routing.
- Built-in thinking mode —
include_thoughts=Truereturns chain-of-reasoning beside the answer. We publish both, so residents can audit the logic. - Long-context with alternating local + global attention — 15-minute audio chunks stitched with a drift-detector pass for hours-long meetings.
- Strict-JSON / response-schema function calling — every output validates against the Pydantic-style schema in
prompts/policy_analysis_v1.md. No prose-parsing fragility. - Mixed-size deployment — Gatekeeper triage on E2B/E4B (edge); deep analysis on 26B-A4B-IT or 31B; safety on ShieldGemma; clustering on EmbeddingGemma.
- Open-weights local fallback —
GOVERNANCE_LLM_BACKEND=huggingfaceswaps every AI Studio call for local HF weights. The same notebook runs end-to-end offline on a single L4 GPU. - Hybrid cloud + edge routing — small files take the cheap E2B route; PDFs ≥1.5 MB and audio >15 min auto-escalate to 31B/26B.
Full feature mapping: ARCHITECTURE.md.
Two entry points — pick whichever fits the next 5 minutes.
Static React app, no API key needed: search + ACS data explorer + Gemma meetings tab.
- Live: https://getcommunityone.github.io/c1_gemma_4_good/
- Local:
npm run install:web && npm run dev - After pipeline §6:
python scripts/colab/export/export_web_demo_index.pyrefreshes meeting/search JSON. - Published website data: https://huggingface.co/datasets/CommunityOne/one-jurisdiction-mapping-analysis
▶ Open run_in_colab.ipynb in Google Colab (step-by-step: JUDGES.md).
- Add Colab Secrets:
GEMINI_API_KEY(free key) andHF_TOKEN(Hugging Face) for Phase 2 video. - Runtime → CPU — run §1 → §5, then §6 Phase 1 (gatekeeper + PDF / Demo 3).
- Runtime → L4 GPU → Restart session → re-run §1 → §5 → complete the checkpoint cells → §6 Phase 2 (ffmpeg + Demo 4).
- Do not use Run all across a runtime switch; the notebook uses interactive dropdown gates between phases.
- Pre-staged demo corpus auto-mounts from Drive (
CommunityOne/hackathons/2026_Gemma_4_Good). - Outputs land in
03_processed_outputs/02_gemma_json/<STATE>/<scope>/<jurisdiction>/…(e.g.AL/county/county_01125/…).
Total time: 45–75 minutes end-to-end on Colab free tier (SCOPE = "fast"). A scripted smoke test without Colab: demos/quickstart.sh.
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
cp .env.example .env # GEMINI_API_KEY; HF_TOKEN for Phase 2 / offline HF
bash scripts/colab/utils/mount_drive.sh # optional: WSL Drive mount
jupyter lab scripts/colab/run_in_colab.ipynbRun §1 → §5 → §6 (two-phase: CPU then GPU). Outputs mirror to Drive under …/03_processed_outputs/02_gemma_json/. Details: scripts/colab/README.md.
Built for communities where privacy is non-negotiable. Three guarantees:
- No login wall for the demo or code. Public Kaggle Notebook, public Colab, CC-BY-4.0.
- No identifying data leaves the local machine in offline mode. The entire pipeline runs against
google/gemma-4-E2B/E4B/26b-a4b-itweights via Hugging Face. Toggle:GOVERNANCE_LLM_BACKEND=huggingface. - ShieldGemma reviews every LLM output before it lands on disk. Demographic enrichment (Demo 5) is opt-in, capped per jurisdiction, and labeled model-perceived, not factual.
c1_gemma_4_good/
├── README.md ← you are here
├── JUDGES.md ← 10-minute judge entry (web + Colab)
├── SUBMISSION.md ← Kaggle writeup (≤ 1,500 words)
├── ARCHITECTURE.md ← system diagram + Gemma 4 feature map
├── RULES_CHECKLIST.md ← every rule mapped to evidence
├── VIDEO_SCRIPT.md ← 3-minute pitch storyboard
├── demos/quickstart.sh ← Gatekeeper smoke test (no Colab)
├── prompts/policy_analysis_v1.md ← deconstruction prompt + JSON schema
├── web/ ← static React UI (GitHub Pages)
├── scripts/colab/run_in_colab.ipynb ← judges run this (Colab or Jupyter)
├── scripts/colab/README.md ← Drive layout, env vars, module map
├── scripts/colab/{engine,demos,triage,…}/ ← pipeline Python (Gemma, gatekeeper, HF)
└── tests/ ← bootstrap + UI tests
Legacy notebook names 02_run_meeting_llm.ipynb / 03_run_meeting_llm.ipynb were renamed to run_in_colab.ipynb.
License: CC-BY-4.0 — per hackathon rules §2.5 (Winner License).
Citation: Ian Ballantyne, Glenn Cameron, María Cruz, Olivier Lacombe, Kristen Quan, and Omar Sanseviero. The Gemma 4 Good Hackathon. https://kaggle.com/competitions/gemma-4-good-hackathon, 2026. Kaggle.
Related: Full production platform (private): open-navigator. This repo is the hackathon-shaped, public, CC-BY-4.0 carve-out.
My wife Renee's favorite song is Defying Gravity. In it, Elphaba says she's "through with playing by the rules of someone else's game."
So am I.
Our kids are on track for a worse financial future than we had — not because the data isn't there, but because it's locked away from the people who need it most. Without facts, we govern on emotion. Without a shared language, we can't even disagree productively.
CommunityOne replaces emotion with evidence. It maps a better path. And it does it for free, for every jurisdiction, in every language people actually speak.
You and I — defying gravity.
