Silent Consensus

Inspiration

We are all roommates. We live together, eat together and plan almost everything in one group chat, and we noticed how often plans quietly fall apart. Nobody fights. Someone just says "idk, maybe" because the place is too expensive, the time doesn't work, or they don't drink, and they don't want to explain why in front of everyone. It's small every time, but those micro-frictions add up, and over time they're how friends drift apart.

It's not just us. 67% of Americans have declined social events because of cost, and 56% never told anyone that money was the reason (CFP Board, 2026). Loneliness affects 1 in 6 people worldwide (WHO, 2025). We wanted a group chat that removes those small tears before they turn into distance.

What it does

Silent Consensus is a group chat with an AI planner, Hush, built in.

You just chat. When a plan starts forming ("concert this month, then food after?"), or a plan starts getting stuck ("maybe not, I'm busy"), Hush notices on its own. You can also tap Plan event.
Hush takes it private. Each person gets a private Hush chat. Hush asks once if they can make it, then only what it actually needs: times, what they're into, roughly where they're coming from, and what's comfortable to spend. It never asks why.
One plan everyone can say yes to. Hush looks at everyone's answers together with names hidden, plus linked calendars (busy/free only). It builds a real plan, with real restaurants and events near the middle of where everyone's coming from, and Resy, OpenTable and ticket links. Everyone confirms it privately before anything is posted.
The group only sees the result. Nobody learns who couldn't afford it, who can't do Friday, or who doesn't drink.
The quiet chip-in. If a plan costs more than someone is comfortable with, friends with room are privately asked if they'd like to quietly help. Payments go through Stripe, and nobody ever knows who gave or who was helped.
It's a real messaging app first. It has accounts, group chats and DMs, replies in threads, reactions, search, and files that are deleted after 7 days (Hush can't read them unless you say yes). There's also "Make Hush yours" personalization and a desktop and phone layout for everything.
Hush for Teams. Work chats track meetings and action items automatically, get a private tone check before a message goes out, and have an admin page.

How we built it

App: Next.js 16 (App Router), React 19, TypeScript and Tailwind, with Prisma on Postgres and row-level security on every table.
AI: every Hush feature goes through one AI client. It uses Meta's Muse Spark first, falls back to Grok, and then to OpenRouter (Jev).
- Every call must return JSON that matches a Zod schema, with automatic repair and fallback.
- Separate AI steps notice plans, run each private conversation, build the plan, pick real places, and check tone.
Real-world data: Stripe Checkout (test mode) for the chip-in, OpenStreetMap for places, and calendar links (ICS) for free/busy.
Hosting: we self-host on a Raspberry Pi 5 behind Cloudflare at silentconsensus.world.
Privacy: it's enforced in code, not just promised. Group-facing data only goes through an allowlisted serializer with tests guarding it. Group-level AI prompts see people as P1, P2… and outputs are checked for names before anyone sees them.

Challenges we ran into

Keeping it private without making it useless. Hush needs everyone's constraints to make a good plan, but it can never reveal them. Hidden names, name-filtered output, and a strict "answers private, outcomes public" rule took a lot of iteration.
Not being annoying. Our first version asked questions in the group chat and asked the same things twice. We redesigned it so everything happens in a private chat: attendance is asked once at the start and confirmation once at the end.
Unreliable AI output. Models sometimes return prose instead of JSON, and API keys failed mid-hackathon. Schema validation, repair retries and a three-provider fallback kept Hush working.
Real places, for free. Public map servers were often overloaded, so we built fallbacks until Hush reliably returned real venues.
Texting compliance. Group texting through Twilio needs A2P 10DLC approval, which takes longer than a hackathon.

Accomplishments that we're proud of

It's live on our own domain, running on a Raspberry Pi in our apartment.
Hush plans a multi-part outing end to end, entirely in private, and comes back with real places and booking links.
A working anonymous chip-in with Stripe, where nobody knows who helped whom.
A 206-point automated check of every screen on desktop and phone, all passing, plus 156 unit tests.

What we learned

Privacy is a product decision, not a checkbox. The hard part is designing flows where nobody ever has to explain themselves.
Structured AI output (schemas, validation, fallbacks) is what makes an AI feature reliable enough to ship.
The best social tool is invisible: people just chat, and the plan shows up.

What's next for Hush

Native apps for Android and iOS. The same Silent Consensus chats, with Hush in your notifications. When a plan needs you, Hush pings you privately, and you answer in a tap without opening the group.
Hush in your native texts. We're spinning Hush off into the group chats and DMs people already use: regular SMS/RCS group texts and 1:1 texts.
- Add Hush's number to any group text and it reads along like any other member.
- Hush checks in with each person in a private 1:1 text, then posts one plan back to the group.
- No new app for your friends to download. The texting pipeline (Twilio Conversations with consent-based opt-in) is already built and waiting on carrier approval.
Friction detection. Gently helping privately when a conversation starts getting tense, before it becomes a fight.
Richer plans and teams. Google Places and Ticketmaster for real venues and events, and Slack/Teams apps plus SSO for Hush for Teams.