Inspiration
Dario Amodei's essay Machines of Loving Grace imagines a future where AI helps strengthen democracy — not by replacing human judgment, but by giving it better infrastructure. That vision was our starting point.We asked: what does a broken democracy actually look like on the ground? The answer isn't dramatic. It's quiet. It's a citizen who doesn't know a consultation is open. It's a councillor who has to guess what their constituents think. It's a person who wants to participate but won't fill out a PDF form at 11pm after a long day.Democratic institutions are strained not because people stopped caring — public trust in UK Parliament sits at just 35% (Ipsos, 2024) — but because the infrastructure for participation hasn't kept up with modern life. The GOV.UK API publishes every open government consultation in real time. Almost no one uses it. The plumbing for democracy exists. It just needs a better front door. REUNITEUS is that front door.
What it does
REUNITEUS directly addresses the Track 4 challenge: helping people participate in democracy, work together across differences, and make collective decisions better. Who we built it for: Citizens who want to be heard but face too much friction to participate — and elected representatives who are making policy decisions without knowing what the public actually thinks.
Part 1 — Global News Bulletin
An AI pipeline reads 30+ international sources daily — BBC, Reuters, Al Jazeera, Fox News, NPR, Le Monde, South China Morning Post, The Africa Report — and delivers the 10 most globally significant stories in plain, accessible language. Sources span every political perspective. No algorithmic personalisation — every user sees the same objective briefing. Informed citizens are the prerequisite for meaningful civic participation.
Part 2 — Civic Engagement Platform
A mobile-first tool that lets any citizen register an opinion in under 10 seconds:
- Tap a topic chip (NHS, Housing, Climate…)
- Swipe a card or tap one of six stance buttons
- Add optional free text — then submit
No account. No email. No friction. Completely anonymous by default. Legislators access a real-time intelligence dashboard: sentiment breakdowns, topic heatmaps, word clouds, and a one-click AI briefing with three concrete recommended actions and an urgency rating — written as if advising a councillor, not feeding a data pipeline. The platform also surfaces live UK government consultations from GOV.UK, a community survey builder for local representatives, and a Government Promise Tracker — colour-coded accountability for pledges made and kept.
How we built it
The architecture is lean by design — powerful enough for real AI workloads, auditable enough for civic tech. Backend: Python 3.13 + Flask, SQLite for zero-infrastructure persistence. The news pipeline runs as: fetcher.py → factfile.py → synthesiser.py → pipeline.py → server.py Incremental refresh means only new articles are processed on subsequent runs, keeping costs to roughly $0.15–$0.40/day. Frontend: Vanilla HTML/CSS/JS — no build tools, no frameworks, one file per page. Civic tech should be auditable and dependency-free. Data Sources: 21 RSS feeds + Guardian API + NYT API for news; GOV.UK Search API (free, no key required) for live consultations.
Challenges we ran into
Prompt reliability across 30+ source types
Getting Claude to produce consistently structured fact-files across a Reuters financial story, an Al Jazeera conflict report, and a Le Monde political profile required careful prompt engineering. Early versions hallucinated significance scores or dropped fields. We solved this with strict JSON schemas, few-shot examples, and retry logic on malformed outputs.
Making participation genuinely frictionless
Usability testing revealed the original design had too many steps. The breakthrough was the bottom sheet — a sliding panel that lets a user react to a news story without ever leaving the news page. Reducing interaction to a single tap transformed the feel from "civic duty" to "natural response." This directly addresses one of Track 4's core challenges: civic participation is down not because people are apathetic, but because participation is designed to be hard.
Legislator dashboard restraint
Presenting sentiment data to elected representatives without it becoming a data dump required genuine design discipline. We stripped out everything that didn't answer "what should I do about this?" — leading with the AI briefing and action recommendations, and burying the raw charts below.
Source diversity vs. quality
Including 30+ global sources sounds clean until you're handling paywalled articles, broken RSS feeds, encoding issues from non-Latin scripts, and duplicate stories from wire services. Deduplication went through several iterations.
What we learned
AI is a UX problem as much as a technical one. The hardest part wasn't getting Claude to summarise news — it was deciding what to surface, in what order, for which audience. A citizen needs plain language and emotional relevance. A legislator needs urgency, specificity, and actionable framing. Same underlying data; two completely different presentation layers.
Civic tech fails on friction, not intent. People want to participate. They won't file a PDF form. The design insight that unlocked this platform was treating civic participation like a mobile-native interaction — swipes, chips, bottom sheets — not a digital version of a paper ballot.
Open government data is dramatically underused. The GOV.UK consultation API is free, documented, and real-time. Almost no civic tools consume it. The infrastructure for democratic participation already exists — it needs a better front door.
Ethics needs to be designed in, not bolted on. Anonymous participation, source diversity, and AI-as-advisor (not AI-as-decider) weren't features we added at the end. They were constraints that shaped every architectural decision from the start.
What's next for ReuniteUS
Geographic filtering — legislators see only opinions from their own constituency Notification system — alert citizens when a consultation matches their stated interests Multi-language support — starting with the top 5 languages spoken in the UK Embeddable widget — any local council deploys the opinion module with a single tag Longitudinal tracking — show how public sentiment on a topic has shifted over months, giving context to political cycles Anonymous civic verification — optional zero-knowledge proof of residency to weight responses from directly affected communities, without compromising anonymity</p>
Log in or sign up for Devpost to join the conversation.