Lynx

The AI Boardroom — Synthesis of Diverse Intelligence.

Lynx is not just a model router; it's a multi-modal orchestration engine designed to consult multiple LLMs simultaneously ("The Boardroom") to find patterns, reach consensus, and identify outliers.

The Vision

Most orchestration tools focus on reducing costs by routing to the cheapest model. Lynx takes a different approach: Intelligence Density.

By streaming responses from multiple frontier models (e.g., GPT-4, Claude 3, Llama 3) in parallel and synthesizing them in real-time, Lynx provides:

• Consensus: High confidence when models agree.
• Nuance: Different perspectives on complex creative or reasoning tasks.
• Reliability: Fallback redundancy and outlier detection.

Architecture

Lynx is built as a TypeScript Monorepo using Turborepo.

Workspace Structure

• apps/api: Node.js backend using Fastify. Handles client connections, manages SSE (Server-Sent Events) streams, and invokes the orchestrator.
• apps/web: React + Vite frontend. Provides the "Boardroom" UI for users to select models and watch parallel streams.
• packages/core: The brain of Lynx. Contains the StreamOrchestrator, ModelStrategy interfaces, and provider logic. It uses advanced async handling to merge distinct model streams into a single unified event stream.
• packages/shared: Shared Zod schemas (LynxRequest, LynxEvent) and TypeScript types used across the entire stack.

Tech Stack

• Language: TypeScript throughout.
• Monorepo: Turborepo, NPM Workspaces.
• Backend: Fastify, Server-Sent Events (SSE).
• Frontend: React, Tailwind CSS, Vite.
• Validation: Zod.
• AI/LLM: Ollama (Local), Mock Providers (Remote simulation).

User Guide

1. Prerequisites (Local Intelligence)

Lynx is configured to use local models via Ollama.

1. Install Ollama.
2. Pull the required models:

   ollama pull gemma3
   ollama pull gpt-oss  # Or any other model you wish to map

3. Start the Ollama server:

   ollama serve

2. Installation & Start

# Install dependencies
npm install

# Start the full stack (API + Web)
npm run dev

• API runs on http://localhost:3000
• Web Interface runs on http://localhost:5173 (typical Vite port)

3. Using the Boardroom UI

1. Open http://localhost:5173 in your browser.
2. Command Center:
   • Enter your prompt (e.g., "Analyze the impact of AI on privacy").
   • Select the models you want to consult (toggle gpt-oss, gemma3, or mocks).
   • Click "SUMMON BOARD".
3. The Workspace:
   • Watch as multiple columns stream text in parallel.
   • Sync Scroll: Scrolling one column will automatically scroll the others for easy line-by-line comparison.
4. Synthesis:
   • The top/highlighted card will show the Aggregated Insight, a synthesized summary of the board's opinions.

Developer Guide

Build

Compile all packages and apps:

npm run build

Verification

To verify the core orchestration logic independently:

# Run the demo script in packages/core
npm run test --workspace=@lynx/core

# Verify API connection to Ollama
npx ts-node apps/api/test-ollama.ts

Detailed Flow

1. Client sends request with prompt and modelIds.
2. API validates request using @lynx/shared schemas.
3. Orchestrator initializes selected providers (Ollama or Mock).
4. Parallel Streaming: Opens streams to all providers concurrently.
5. Merging: Chunks are yielded instantly as they arrive from any provider via Promise.race.
6. Synthesis: Aggregates results after streams finish for a final summary.
7. Client receives a single server-sent event stream updating the UI in real-time.