The Spark
The inspiration for A²MP came from a place we've all been: a pointless meeting.
I was in a weekly team check-in meeting during an internship with two of my teammates. About ten minutes in, one of them sighed and said the one thing every professional feels: "This really could have been an email." We all laughed, but it sparked a real discussion.
The problem wasn't that the topic was useless; we did need to align, but the format was broken. That's when I remembered an old episode of Black Mirror. In the show, a dating app runs thousands of complex simulations to determine a couple's perfect match, all before they even meet.
The idea clicked!
What if we applied that same simulation concept to the workplace? What if, instead of running simulations to find a "perfect match," we could run a simulation to find a "perfect decision"?
Could we build a tool where everyone submits their detailed, asynchronous input, and AI agents representing our unique goals hold the meeting for us? This project is the first step toward that answer.
The Vision
The idea consumed me. I imagined a world where:
- A product manager in San Francisco could have their AI persona debate feature priorities with an engineering lead's persona in Berlin
- Budget discussions could happen between AI representatives of different departments, working through constraints and trade-offs
- Strategic decisions could emerge from AI-mediated conversations that considered everyone's input, even when they couldn't all be in the same room
This wasn't about replacing human judgment – it was about augmenting human collaboration.
The Technical Challenge
Persona Generation
How do you capture someone's perspective, priorities, and communication style in an AI agent? I developed a system that analyzes participant inputs to generate unique personas with:
- Identity and role descriptions
- Specific objectives and constraints
- Natural communication patterns
- Decision-making frameworks
Real-Time Conversation Engine
The heart of A²MP is its conversation engine that:
- Orchestrates turn-taking between AI personas
- Detects when conversations are getting stuck in loops
- Manages the flow from discussion to consensus
- Pauses for human intervention when needed
Smart Moderation
I implemented an AI moderator that uses sophisticated logic to decide who speaks next:
- Prioritizes direct responses to questions
- Ensures everyone gets a chance to contribute
- Alternates speakers to maintain natural flow
- Recognizes when human guidance is needed
The Technology Stack
I chose modern, robust technologies to bring this vision to life:
- Backend: Node.js with Express and TypeScript for type-safe, scalable APIs
- Frontend: Next.js 14 with React for a responsive, real-time user interface
- Real-time Communication: Socket.IO for seamless WebSocket connections
- AI Integration: Google Gemini 2.5 Flash for powerful language understanding
- Database: SQLite for reliable, file-based persistence
- UI Components: Tailwind CSS with custom components for a polished experience
The Challenges
Rate Limiting & Quota Management
AI APIs have strict rate limits, so I implemented a dual-key system that separates persona responses from moderation decisions, effectively doubling our capacity to 500+ requests per day.
Repetition Detection
Early versions would sometimes get stuck in conversational loops. I built sophisticated pattern detection that identifies when personas are repeating themselves and automatically pauses for human intervention.
Real-Time Synchronization
Keeping multiple clients synchronized during live conversations required careful WebSocket management and state synchronization across the entire application.
Natural Turn-Taking
Creating conversations that feel natural rather than robotic required implementing priority logic that mimics human conversation patterns – responding to direct questions, alternating speakers, and managing interruptions gracefully.
The Lessons
Technical Insights
- AI Orchestration: Managing multiple AI agents in real-time conversations is as much about conversation design as it is about technical implementation
- WebSocket Architecture: Real-time collaboration requires careful consideration of connection management, room-based broadcasting, and graceful error handling
- Persona Design: The quality of AI conversations depends heavily on how well you capture and model individual perspectives and communication styles
Product Insights
- Asynchronous Collaboration: There's enormous untapped potential in tools that work across time zones and schedules
- AI Augmentation: The most powerful AI applications don't replace humans – they amplify human capabilities and remove friction from collaboration
- Trust & Transparency: Users need to understand and trust how their personas are representing them, requiring clear visualization of AI decision-making
The Impact
A²MP transforms how distributed teams make decisions. Instead of:
- Scheduling conflicts across time zones
- Rushed decisions in time-constrained meetings
- Quiet voices getting overshadowed
- "This could have been an email" frustration
We enable:
- Asynchronous consensus-building that respects everyone's schedule
- Thoughtful AI-mediated discussions that consider all perspectives
- Automatic documentation with complete transcripts and decision summaries
- Equitable participation where every voice is heard and represented
The Future
This project opened my eyes to the transformative potential of AI in collaborative work. A²MP is just the beginning – imagine AI personas that:
- Learn and adapt from past decisions
- Integrate with existing business tools and data
- Derive rich context from productivity platforms like Notion and Obsidian
- Handle complex multi-stakeholder negotiations
- Provide insights into team dynamics and decision patterns
The future of work isn't just about working from anywhere – it's about working anytime, with AI as our collaborative partner.