Sidebar | Devpost

Landing Page
Mid debate in session, with visual personality vector embeddings for distinct personalities
Final Verdict

Inspiration

We make a lot of decisions we're not equipped for, alone, at 2am, with a group chat that will not give you a straight answer. Sometimes you want the cold hard truth. Sometimes you want someone to hype you up. Sidebar is what happens when you build a panel of AI voices that cover both, and then make them actually argue about it in front of you instead of just picking one and hoping for the best.

What it does

Give Sidebar your dilemma and it puts together a panel of three to six AI personas, cast specifically to disagree with each other. They open with their own take, go a round of direct rebuttal, deliver closing pitches, and then a Chair rules on it. Real verdict, a plan you can actually use, and the best version of the argument that lost, stated fairly instead of quietly dropped. No more asking one chatbot and getting told you were right all along.

How we built it

Gemini powers the personas and the Chair. A Hono backend runs the whole deliberation and streams it live over SSE. The frontend is Next.js. Casting the panel is a MongoDB Atlas and Voyage embeddings job, vector search to find relevant personas, then a diversity pass so you don't end up with five people who all studied the same major in undergrad. A shared Zod contract keeps the frontend and backend from lying to each other about what a verdict actually looks like.

Challenges we ran into

Getting AI personas to actually disagree instead of agreeing with everything, models are professional yes-men by default
Casting a panel that's diverse and not just different names with the same opinion
Keeping a live, multi-agent argument perfectly replayable after the fact

Accomplishments that we're proud of

A verdict that feels like an actual ruling, not a vibes-based summary
Personas with enough personality that we started taking sides ourselves
The whole deliberation streams live and replays exactly as it happened

What we learned

Getting AI to disagree on purpose is a lot harder than getting it to agree, since models default to being agreeable no matter what. We also learned a ton about casting for genuine diversity instead of just relevance, and that a shared schema between frontend and backend saves you from yourself when everyone's moving fast.

What's next for Sidebar

Grow the persona library way past our starting set, give the panel real tools like live search, and let people jump back into a specific argument instead of just accepting the ruling and walking away.