BizRoom.ai

version	2.0.0
created	2026-03-11 22:00
updated	2026-03-16 09:00

BizRoom.ai

Your AI Executive Team — Meet with 6 AI C-Suite executives in a 3D virtual boardroom.

Microsoft AI Dev Days Hackathon submission. Categories: Grand Prize (AI Applications & Agents) + Best Multi-Agent System

Hero Technologies

BizRoom.ai is built on four Microsoft/Azure hero technologies — each with direct code evidence.

Technology	Role in BizRoom.ai	Source File
Azure AI Foundry Model Router	Single endpoint auto-routes tasks to optimal models (quality/cost/balanced)	`backend/src/services/ModelRouter.ts`
Microsoft Agent Framework	EventEmitter-based TurnManager state machine, P0–P4 priority queue, A2A routing	`backend/src/orchestrator/TurnManager.ts`
Azure MCP Server	JSON-RPC 2.0 endpoint exposing 3 tools (Excel, PPT, meeting summary) to AI hosts	`backend/src/functions/mcp.ts`
GitHub Copilot Agent Mode	Full-stack TDD development — subagent-driven-development pattern throughout	Entire codebase (TypeScript strict, vitest TDD)

What is BizRoom.ai?

The Problem

Every business decision requires multiple expert perspectives — finance, marketing, legal, technology, design, and operations. But solo founders, startups, and SMBs cannot afford a full C-Suite.

The result: critical decisions made with blind spots.

The Solution

BizRoom.ai puts 6 AI C-Suite executives in a real-time voice meeting with you. You speak — they listen, debate, and advise. The meeting auto-generates PPT slides, Excel reports, and Planner action items in OneDrive.

You (voice/text)
       │
       ▼ Azure Functions (20 endpoints)
  TurnManager — P0–P4 priority state machine
       │
       ▼ TopicClassifier
  Relevant agents selected
       │
  ┌────┴────────────────────────────────────────┐
  │                                             │
  ▼                  ▼                          ▼
Hudson (COO)     Amelia (CFO)           Bradley (CLO)
Kelvin (CTO)      Yusef (CMO)            Jonas (CDO)
  │
  ▼ StructuredAgentOutput
  { speech, key_points, mention, visual_hint }
       │
       ├──▶ HoloMonitor3D  (key_points per avatar)
       ├──▶ A2A mention routing (follow-up turns)
       │
       ▼ Sophia Pipeline
  visual_hint → BigScreenRenderData
       │
       ▼ Azure SignalR Service → React 19 Frontend (3D)

Screenshots

Demo Video: YouTube

Lobby — Brand Memory Setup

3D Boardroom — Full Executive Team

Live Meeting — Real-Time Agent Conversation

Agent Discussion — Multi-Agent Debate

BigScreen — Sophia Real-Time Visualization

Architecture Overview

Orchestration Pipeline

User voice/text
    │
    ▼ POST /api/message  or  WebSocket (GPT Realtime 1.5)
    │
    ▼ TurnManager.handleUserInput()
      ├── P0: Interrupt (user takes floor immediately)
      ├── P1: Direct mention (@agent)
      ├── P2: Topic-classified agents (TopicClassifier)
      ├── P3: A2A follow-up (mention routing)
      └── P4: Auto-mode rotation
    │
    ▼ AgentFactory.invokeAgent()
      └── ModelRouter
            ├── Azure AI Foundry Model Router  (primary)
            └── Anthropic / OpenAI             (fallback)
    │
    ▼ ResponseParser → StructuredAgentOutput
      { speech, key_points, mention, visual_hint }
    │
    ├──▶ VoiceLiveOrchestrator.agentDone()
    │     ├── sophiaAgent.addToBuffer()      (visual queue)
    │     ├── SignalR: monitorUpdate         (HoloMonitor3D)
    │     ├── TurnManager.handleMention()   (A2A routing)
    │     └── TurnManager.onAgentDone()     (advance turn)
    │
    └──▶ Sophia Pipeline (FIFO queue)
          visual_hint → callSophiaVisualGPT()
          → BigScreenRenderData
          → SignalR: bigScreenUpdate → BigScreenRenderer (3D)

Model Routing (ModelRouter.ts)

Task Type	Primary	Fallback Provider	Fallback Model	Temp
`agent-response`	Foundry model-router	Anthropic	claude-opus-4-6-20250929	0.5
`minutes`	Foundry model-router	Anthropic	claude-opus-4-6-20250929	0.4
`summary`	Foundry model-router	Anthropic	claude-opus-4-6-20250929	0.4
`visual-gen`	Foundry model-router	Anthropic	claude-sonnet-4-6-20250514	0.2
`visual-gen-fast`	Foundry model-router	Anthropic	claude-haiku-4-5-20251001	0.2
`parse-fallback`	Foundry model-router	Anthropic	claude-haiku-4-5-20251001	0.1
`realtime-voice`	Foundry model-router	OpenAI	gpt-realtime-1.5	0.6
`deep-analysis`	Foundry model-router	OpenAI	o3	0.1

Switching to Azure AI Foundry: Set one environment variable — AZURE_FOUNDRY_MODEL_ROUTER_ENDPOINT — and all task routing switches to the Foundry endpoint automatically. Zero code changes required.

STT: Web Speech API (primary) + Whisper-1 fallback via /api/voice/transcribe.

MCP Server (mcp.ts)

BizRoom exposes a fully spec-compliant MCP server at GET|POST /api/mcp:

Tool	Description
`bizroom_generate_excel`	Generate Excel budget report (.xlsx)
`bizroom_generate_ppt`	Generate PowerPoint from meeting minutes (.pptx)
`bizroom_meeting_summary`	Real-time summary of active meeting room (by `roomId`)

Protocol: JSON-RPC 2.0, MCP version 2025-06-18. Compatible with any MCP-capable AI host (GitHub Copilot, Claude Desktop, etc.).

AI Agent Lineup

Role	Name	Inspiration	Specialty
COO	Hudson	Judson Althoff	Operations, execution, cross-functional alignment
CFO	Amelia	Amy Hood	Financial modeling, KPIs, budget analysis
CMO	Yusef	Yusuf Mehdi	Brand strategy, GTM, growth marketing
CTO	Kelvin	Kevin Scott	Technical architecture, engineering, scalability
CDO	Jonas	Jon Friedman	Product design, UX/CX, design systems
CLO	Bradley	Brad Smith	Legal risk, compliance, contracts, IP
(Support)	Sophia	Analysis Agent	Real-time visualization + meeting minutes

Sophia is not a C-Suite member. She operates exclusively via the internal VoiceLiveOrchestrator event pipeline — no HTTP API endpoint.

Agent-to-Agent (A2A) Communication

Every agent response includes a StructuredAgentOutput:

{ speech: string, key_points: string[], mention: Mention | null, visual_hint: VisualHint | null }

When an agent sets the mention field, TurnManager routes the next turn to the mentioned agent — creating genuine multi-agent debate rather than scripted dialogue. Follow-up rounds are tracked and capped to prevent runaway chains.

Core Features

1. Real-Time Voice Meeting (Live Mode)

Push-to-Talk (PTT) captures audio via Web Speech API. Transcript is sent to TurnManager, agents respond sequentially via GPT Realtime 1.5 WebSocket. Ready Player Me avatars animate with viseme lip-sync on every response.

2. Autonomous Discussion (Auto Mode)

The board discusses a given agenda topic without user intervention. TurnManager rotates agents by topic relevance, enforcing a maximum of 2 consecutive AI responses before surfacing the user for input.

3. Direct Message Mode (DM Mode)

1:1 conversation with a specific agent. Other agents are silent. Useful for deep dives into a single domain (e.g., legal review with Bradley, financial projection with Amelia).

4. Sophia Real-Time Visualization

Every agent response may include a visual_hint. Sophia buffers hints in a FIFO queue and generates BigScreenRenderData via Claude Sonnet 4.6. Results broadcast via SignalR bigScreenUpdate and rendered on the 3D BigScreen in the boardroom.

Supported visual types: bar_chart, pie_chart, line_chart, swot_matrix, market_map, kpi_cards, summary_card, checklist, timeline, table.

5. Automated Meeting Artifacts

On POST /api/meeting/end:

COO Hudson delivers structured closing remarks
Sophia generates meeting minutes via Claude Opus 4.6
pptxgenjs generates PowerPoint + exceljs generates Excel
Files uploaded to OneDrive via Microsoft Graph API
Action items created as Microsoft Planner tasks
artifactsReady SignalR event triggers the frontend download UI

6. MCP Integration

External AI agents (GitHub Copilot, Claude Desktop) can call BizRoom tools via the MCP server at /api/mcp — generating Excel reports, PPT decks, or querying live meeting summaries in real time.

Tech Stack

Frontend

Technology	Version	Role
React	19.x	UI framework (Concurrent Mode)
TypeScript	5.x strict	Type safety across all components
Tailwind CSS	4.x	Utility-first styling (`@tailwindcss/vite`)
Vite	7.x	Build tooling + HMR dev server
React Three Fiber + @react-three/drei	9.x / 10.x	3D virtual boardroom
Three.js	0.170	WebGL renderer
@microsoft/signalr	10.x	Real-time WebSocket client
Ready Player Me	SDK	AI agent 3D avatars + viseme lip-sync

Backend

Technology	Version	Role
Azure Functions v4	Node 20	Serverless API — 20 HTTP endpoints
TypeScript	5.x	Strict mode throughout
OpenAI SDK	4.x	GPT Realtime 1.5 WebSocket voice + Whisper-1 STT
Anthropic SDK	0.78+	Claude Opus / Sonnet / Haiku for text generation
pptxgenjs	3.x	PowerPoint generation
exceljs	4.x	Excel generation
ws	8.x	WebSocket relay for GPT Realtime
@modelcontextprotocol/sdk	latest	MCP tool type definitions
vitest	1.x	Unit testing (TDD throughout)

Infrastructure

Service	Tier	Role
Azure Functions v4	Consumption	Backend API (20 endpoints)
Azure SignalR Service	Serverless / Premium_P1	Real-time WebSocket hub ("default")
Azure Static Web Apps	Standard	Frontend hosting + CDN + API proxy
Azure Cosmos DB	Serverless	Conversation history, room sessions, user profiles
Azure Blob Storage	LRS	Artifact storage (PPT, Excel, meeting minutes)
Azure AI Foundry	Pay-per-token	Model Router — primary AI inference endpoint
Microsoft Graph API	—	OneDrive file upload, Planner task creation

Getting Started

Prerequisites

Node.js 20+
Azure Functions Core Tools v4
Azure CLI (for deployment)
API keys: ANTHROPIC_API_KEY, OPENAI_API_KEY
Azure resources: SignalR Service, Cosmos DB, Blob Storage, Static Web Apps

Local Development

# 1. Clone the repository
git clone https://github.com/your-org/bizroom.git
cd bizroom

# 2. Install dependencies
cd backend && npm install
cd ../frontend && npm install

# 3. Configure backend environment
cp backend/local.settings.json.example backend/local.settings.json
# Fill in: ANTHROPIC_API_KEY, OPENAI_API_KEY, SignalR connection string,
#          Cosmos DB connection, Blob Storage connection

# 4. Start backend (Azure Functions local runtime)
cd backend && npm run start

# 5. Start frontend (Vite dev server)
cd frontend && npm run dev

# 6. Open http://localhost:5173

Key Environment Variables

Variable	Required	Description
`ANTHROPIC_API_KEY`	Yes	Anthropic API (Claude Opus / Sonnet / Haiku)
`OPENAI_API_KEY`	Yes	OpenAI (GPT Realtime 1.5, Whisper-1)
`AZURE_SIGNALR_CONNECTION_STRING`	Yes	Azure SignalR Service
`COSMOS_DB_CONNECTION_STRING`	Yes	Azure Cosmos DB
`AZURE_STORAGE_CONNECTION_STRING`	Yes	Azure Blob Storage (artifacts)
`AZURE_FOUNDRY_MODEL_ROUTER_ENDPOINT`	Optional	Enables Azure AI Foundry as primary model router
`AZURE_FOUNDRY_API_KEY`	Optional	API key for Foundry endpoint
`MICROSOFT_GRAPH_CLIENT_ID`	Optional	Microsoft Graph (OneDrive + Planner integration)

Running Tests

cd backend
npm run test         # vitest unit tests
npm run test:watch   # watch mode

Judging Criteria Alignment

1. Technical Implementation (20%)

Criterion	BizRoom.ai Implementation
Code quality	TypeScript 5 strict mode — 100% frontend and backend
Testing	TDD with vitest — TurnManager, ResponseParser, TopicClassifier unit tests
Architecture	Clean separation: orchestrator / agents / services / functions / shared types
Scale	20 Azure Functions endpoints, 6 agent system prompts, multi-provider model routing

2. Agentic Design (20%)

Criterion	BizRoom.ai Implementation
Agent autonomy	TurnManager P0–P4 priority queue; agents self-select follow-up via `mention`
A2A comms	`mention` in StructuredAgentOutput → TurnManager routes next agent turn
State machine	EventEmitter: idle → buffering → processing → speaking → idle
Debate control	Max 2 consecutive AI responses; follow-up rounds tracked; human callout enforced

3. Real-World Impact (20%)

Criterion	BizRoom.ai Implementation
Target problem	Solo founders and SMBs lack C-Suite expertise for critical decisions
Market size	Global SMB SaaS market: $1.5T+ (Gartner 2024)
Measurable value	Replaces $50K–$200K/year executive advisory cost with always-on AI board
Tangible output	PPT, Excel, meeting minutes, Planner tasks auto-generated per meeting

4. UX / Presentation (20%)

Criterion	BizRoom.ai Implementation
Immersive UX	React Three Fiber 3D boardroom — avatars, BigScreen, HoloMonitor, ambient lighting
Voice-first	PTT with GPT Realtime 1.5 WebSocket, viseme lip-sync on 3D avatars
Real-time visuals	Sophia renders 10 visual types on BigScreen as the conversation progresses
Zero friction	Web-based, no install. Enter agenda → meeting starts in seconds

5. Category Compliance (20%)

Category	Compliance
Grand Prize: AI Applications	Full-stack AI application — voice, text, 3D, artifacts, multi-agent orchestration
Best Multi-Agent System	6 C-Suite agents + Sophia, TurnManager, A2A mention routing, topic classification
Hero: Azure AI Foundry	ModelRouter with Foundry endpoint (`AZURE_FOUNDRY_MODEL_ROUTER_ENDPOINT`)
Hero: Microsoft Agent Framework	TurnManager = custom agent framework (EventEmitter, priority queue, A2A routing)
Hero: Azure MCP	`/api/mcp` — JSON-RPC 2.0, MCP 2025-06-18, 3 tools, CORS-ready
Hero: GitHub Copilot Agent Mode	Entire codebase developed with Copilot Agent Mode (TDD, subagent-driven-dev)

Azure AI Foundry — One-Variable Switch

The codebase is production-ready for Azure AI Foundry today. A single environment variable switches all model routing from the current multi-provider fallback to the Foundry Model Router:

// backend/src/services/ModelRouter.ts
if (process.env.AZURE_FOUNDRY_MODEL_ROUTER_ENDPOINT) {
  return "foundry"; // → single endpoint, automatic model selection
}

Expected improvements once Foundry quota is provisioned:

Metric	Current (multi-provider)	With Foundry (projected)	Improvement
Agent response TTFT	~400–600 ms	~150–250 ms	~60% faster
Sophia visualization latency	~800–1,200 ms	~400–600 ms	~50% faster
Voice round-trip latency	~300–500 ms	~150–250 ms	~50% faster
Meeting minutes generation	~3–5 s	~1.5–2.5 s	~50% faster
Auth complexity	3 providers × API key	Azure RBAC single identity	Unified

Project Structure

BizRoom/
├── frontend/src/
│   ├── components/
│   │   ├── meeting3d/    — MeetingRoom3D, BigScreenRenderer, SophiaBlob3D, HoloMonitor3D
│   │   ├── chat/         — ChatRoom, MessageBubble, SophiaMessage, TypingIndicator
│   │   ├── input/        — PushToTalk, MicToggle, InputArea, QuickActions
│   │   └── meeting/      — MeetingBanner, ChairmanControls, DmAgentPicker
│   ├── hooks/            — useSignalR, useVoiceLive, usePushToTalk, useAgentAudio, useViseme
│   └── constants/        — strings.ts (i18n-ready), agentVoices.ts, brandPresets.ts
│
├── backend/src/
│   ├── functions/        — 20 Azure Functions HTTP endpoints (+ mcp.ts)
│   ├── orchestrator/     — TurnManager, VoiceLiveOrchestrator, ContextBroker,
│   │                       ResponseParser, TopicClassifier, SnippetManager
│   ├── agents/           — AgentFactory, SophiaAgent, agentConfigs.ts, prompts/
│   ├── services/         — ModelRouter, VoiceLiveSessionManager, SignalRService,
│   │                       ArtifactGenerator, ArtifactService, GraphService
│   ├── plugins/          — ExcelPlugin, MeetingMinutesPlugin
│   └── constants/        — turnConfig.ts, responseSchema.ts, agentVoices.ts
│
├── shared/
│   └── types.ts          — Single Source of Truth: all shared types (frontend + backend)
│
└── docs/                 — Architecture, Tech Spec, Agent Design, Design System, PRD

License

This software is proprietary and confidential. Unauthorized copying, modification, distribution, or use of this software, via any medium, is strictly prohibited.

Built for the Microsoft AI Dev Days Hackathon — March 2026.

Name		Name	Last commit message	Last commit date
E2ETest		E2ETest
backend		backend
docs		docs
frontend		frontend
infra		infra
shared		shared
.env.example		.env.example
.gitignore		.gitignore
.prettierignore		.prettierignore
.prettierrc		.prettierrc
CLAUDE.md		CLAUDE.md
README.md		README.md
eslint.config.js		eslint.config.js
package-lock.json		package-lock.json
package.json		package.json

Latest commit

History

Folders and files

Repository files navigation

BizRoom.ai

Hero Technologies

What is BizRoom.ai?

The Problem

The Solution

Screenshots

Lobby — Brand Memory Setup

3D Boardroom — Full Executive Team

Live Meeting — Real-Time Agent Conversation

Agent Discussion — Multi-Agent Debate

BigScreen — Sophia Real-Time Visualization

Architecture Overview

Orchestration Pipeline

Model Routing (ModelRouter.ts)

MCP Server (mcp.ts)

AI Agent Lineup

Agent-to-Agent (A2A) Communication

Core Features

1. Real-Time Voice Meeting (Live Mode)

2. Autonomous Discussion (Auto Mode)

3. Direct Message Mode (DM Mode)

4. Sophia Real-Time Visualization

5. Automated Meeting Artifacts

6. MCP Integration

Tech Stack

Frontend

Backend

Infrastructure

Getting Started

Prerequisites

Local Development

Key Environment Variables

Running Tests

Judging Criteria Alignment

1. Technical Implementation (20%)

2. Agentic Design (20%)

3. Real-World Impact (20%)

4. UX / Presentation (20%)

5. Category Compliance (20%)

Azure AI Foundry — One-Variable Switch

Project Structure

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages