Skip to content

DEV Community

Krushna Kodgirwar

Posted on Oct 2

AI Boardroom: Local Multi-Agent Business Decision Simulator

#devchallenge #weekendchallenge #hf26challenge #ai

*This is a submission for the Hacktoberfest Weekend challenge: Build for a Friend*

AI Boardroom — A Multi-Agent AI Decision Simulator Built for a Friend

What I Built

I built AI Boardroom, an open-source multi-agent decision-making simulator designed specifically for a close friend who was struggling to evaluate complex business decisions.

Instead of asking a single AI model for advice and receiving a surface-level, biased response, AI Boardroom creates a virtual executive boardroom. It spins up specialized AI agents—such as a cautious CFO, a growth-focused CMO, a risk-averse Legal Advisor, and a Pragmatic Engineer—who analyze the exact same problem simultaneously, challenge each other's assumptions, and generate a balanced, multi-perspective recommendation.

When I ran his latest startup proposal through the local board in front of him, his initial reaction was:

"Seeing the CFO agent immediately call out my unit economics while the CMO agent argued for brand exposure gave me more clarity in 30 seconds than days of overthinking."

Key Features

Multi-Agent Simulation: Orchestrates multiple distinct personas with tailored prompts and temperature parameters.
Autonomous & Manual Collaboration: Observe agents analyze, critique, and build on each other's points in real-time.
Structured Final Synthesis: Generates an aggregated consensus, highlighted risks, and actionable next steps.
100% Local Open-Weight Execution: Operates completely offline with zero API costs using open-source LLMs.

Code

krushnakodgirwar / AI-Boardroom

AI-powered multi-agent business decision simulator with executive agents and an AI CEO decision system.

🏛️ AI Boardroom OS

### Where AI Agents Think Together

A Multi-Agent AI Business Decision Simulator
Analyze business challenges through specialized AI executives, collaborative reasoning, and structured decision-making

🌐 Live Application

🔗 Live Website: https://ai-boardroom-xexp.onrender.com

🔗 GitHub Repository: https://github.com/krushnakodgirwar/AI-Boardroom

Deployment note: The frontend is hosted on Render. Full AI analysis requires the FastAPI backend to be running and the frontend to be configured with a reachable backend URL.

📑 Table of Contents

📖 Explore the README sections

✨ Introduction

AI Boardroom OS is a multi-agent AI business decision simulator that explores how specialized…

The repository contains the complete source code, including the FastAPI backend, prompt orchestration pipelines, agent persona configs, and interactive frontend interface.

How I Built It

I engineered the entire stack from scratch to run efficiently on consumer hardware without reliance on closed third-party APIs:

1. System Architecture & Tech Stack

Core Intelligence: Open-weight Qwen2.5-3B-Instruct model.
Model Optimization: Loaded via PyTorch using 4-bit NF4 quantization (bitsandbytes), drastically reducing GPU memory overhead so it runs smoothly on standard hardware.
Backend API: FastAPI & Uvicorn handling asynchronous multi-agent prompt pipelines and stream responses.
Frontend UI: Built with HTML5, CSS3, and modern JavaScript (ES6) for clear visual separation of agent discussions.
Tunneling & Testing: Integrated Cloudflare Tunnels for secure local-to-web testing and live previews.

                     USER INPUT
                         │
                         ▼
                Business Problem Context
                         │
                         ▼
              AI BOARDROOM ORCHESTRATOR
                         │
        ┌────────────────┼────────────────┐
        │                │                │
        ▼                ▼                ▼
   Agent 1: CFO     Agent 2: CMO    Agent 3: Tech Lead
(Financial Risk)  (User Growth)    (Feasibility)
        │                │                │
        └────────────────┼────────────────┘
                         │
                         ▼
              Agent Discussion & Review
                         │
                         ▼
             Structured Final Consensus

Why Does Open Innovation Matter?

Building AI Boardroom on open-source technology made critical features possible that closed proprietary APIs couldn't offer:

Unrestricted Persona Fine-Tuning & Prompt Depth: Proprietary APIs often apply heavy global system instruction layers that dull distinct agent personas. Using an open-weight model like Qwen2.5-3B-Instruct allowed me to fine-tune exact agent behavior, context lengths, and temperature parameters per role without safety filter over-triggering.
Absolute Confidentiality: Business ideas, pitch decks, and internal financials are highly sensitive. Keeping inference 100% local ensures zero data is logged or sent to cloud servers to train external models.
Zero Token Cost for Complex Multi-Turn Loops: Multi-agent discussions consume thousands of tokens in context-swapping per run. Running locally eliminated per-token API costs completely, enabling infinite iterations for free.

Challenges I Solved

Context Swapping Latency: Running multi-agent prompts sequentially on a single GPU can slow down response times. I optimized memory consumption with 4-bit NF4 quantization and streamlined prompt templates to keep response latency low.
Preventing Consensus Bias: Early iterations saw agents blindly agreeing with each other. I tuned the temperature settings individually per agent (e.g., lower temperature for CFO, higher for CMO) and added explicit counter-argument system constraints to force realistic boardroom debate.

Prize Categories

Best Use of Gemma / Open-Weight Models

Top comments (0)

Subscribe