Kapwa

The Problem

AI chatbots give you one voice, one perspective, one personality. But real decisions benefit from multiple viewpoints. What if you could sit down with a board of advisors — any time, on any question — each with different expertise, reasoning styles, and perspectives?

That's what Kapwa gives you.

Our Approach

We built a platform where users convene an AI advisory board — selecting up to three personas from a rich library of 288+, inspired by historical figures, domain experts, and creative thinkers — who engage simultaneously, like an actual board meeting.

The key technical challenge wasn't generating responses. It was orchestration: how do you coordinate multiple AI personas so their responses build on each other, maintain coherent context, and produce something more valuable than three separate conversations?

Key Features

AI Advisory Board (Persona Mode)

Users convene a board of up to three AI personas from a rich library of 288+, each defined with distinct expertise, personality traits, communication style, and reasoning approach. The board engages with the user simultaneously — like sitting in an actual board meeting. Users can search the library or generate custom personas on demand.

Symphony Mode

A separate core innovation. Rather than assembling a board of personas, Symphony Mode sends the user's query to three models reasoning in parallel. An orchestrator evaluates all responses and selects the best one — giving the user the depth of multi-model reasoning with the simplicity of a single chat.

Orchestration Engine

Under the hood, a streaming ensemble manager powers both modes:

• Persona Mode: Routes questions to the selected board of personas, maintains each persona's distinct voice, and presents their perspectives in a single chat thread
• Symphony Mode: Dispatches the query to three models in parallel, surfaces individual model responses in a "Thinking" tab, and presents the orchestrator's selected best response in the Chat tab
• Manages conversation context and memory across sessions in both modes
• Handles real-time streaming of concurrent responses

Semantic Memory

Every conversation is embedded and stored in a vector database (pgvector). The system can retrieve relevant context from past conversations, giving personas long-term memory and the ability to reference previous discussions.

Real-Time Streaming

Responses stream in real-time — users see each persona's thinking as it happens, not after a long wait. The streaming architecture handles multiple concurrent persona responses without blocking.

Architecture

The system is built in four layers:

Frontend: React + TypeScript SPA with real-time streaming UI, conversation management, and advisor selection.

Backend: Express server handling API routing, session management, streaming orchestration, and persona management.

AI Layer: Google Gemini models with LangChain for structured workflows. Custom ensemble manager handles multi-persona coordination and response streaming.

Data Layer: Supabase (PostgreSQL) for structured data, pgvector for semantic search and conversation memory, custom embedding pipeline for context retrieval.

Results

• 288+ AI personas available
• Real-time multi-persona streaming conversations
• Sub-second response initiation
• Persistent conversation history with semantic search
• Production-deployed on Railway with continuous delivery