Inspiration
Every enterprise product decision — a pricing change, a feature tradeoff, a compliance question — gets debated in a real room by real people with competing incentives. Finance cares about margin. Legal cares about risk. Engineering cares about technical debt. Most AI tools built for this kind of decision collapse all of that into a single model giving a single, confident answer — which is the opposite of how good decisions actually get made. We wanted to build the thing that argues with itself instead of agreeing with you.
What it does
The Boardroom simulates a live cross-functional review meeting. You submit a real decision, and five persona-driven AI agents — Product, Finance, Legal/Compliance, Customer Success, and Engineering — debate it out loud, in distinct voices, each arguing from its own incentive structure. They open with individual positions, cross-examine each other on points of disagreement, and converge on a recommendation — with any unresolved dissent explicitly logged rather than smoothed over. The output is a structured decision memo: recommendation, key tradeoffs, and named dissent, not just a transcript.
How we built it
The core is a LangGraph state machine orchestrating a three-round debate: parallel opening statements, a targeted rebuttal round where the moderator forces disagreeing agents to respond directly to each other, and a convergence round that synthesizes the final memo. Each persona is grounded in a distinct incentive prompt rather than a generic personality — the Legal/Compliance agent is additionally grounded via a lightweight RAG layer against policy-style documents, so its objections are drawn from real constraints instead of invented ones.
Voice is handled by ElevenLabs, with a distinct voice ID per persona streamed per-turn over WebSockets, so the frontend can play audio as each agent speaks rather than waiting for the full debate to resolve. The UI is a green-on-black terminal aesthetic — deliberately built to feel like a system, not a chatbot — moving through four screens: a landing/pitch screen, a mission-briefing input screen, the live multi-agent debate view, and the decision memo output. Frontend is vanilla HTML/CSS/JS, chosen specifically to keep iteration fast during a 24-hour build without framework overhead.
Challenges we ran into
Getting agents to genuinely disagree rather than politely converge is harder than it sounds — early prompt drafts had every persona hedging toward consensus regardless of their stated incentive. Solving this meant explicitly modeling the rebuttal round as its own graph stage with a moderator that selects real points of tension, rather than trusting a single-pass prompt to produce adversarial reasoning on its own. We also had to think carefully about legibility under real demo conditions — a dark terminal aesthetic that looks great on a laptop can lose all contrast under stage lighting, so idle-vs-speaking states and font sizing were designed with a projector environment in mind from the start, not as an afterthought.
What's next for The Boardroom
If shortlisted, the 24-hour build focuses on turning the full three-round debate loop into a working, streaming, voiced experience end-to-end, with the decision memo as the closing beat of the live Main Stage demo.
Built With
- azure-openai
- claude
- css3
- elevenlabs
- fastapi
- figma
- ggraph
- html5
- javascript
- langchain
- multi-agent-systems
- python
- rag
- websockets
