🛰️ Agent Flow: The Autonomous Meeting Swarm

Eliminating Ambiguity. Shipping Clarity. End-to-End.

Agent Flow is a high-fidelity, autonomous multi-agent platform designed to audit live corporate meetings, neutralize ambiguity in real-time, and generate actionable, high-grade analytical reports without human intervention.

---

🏛️ System Architecture

Agent Flow operates on a Distributed Agentic Architecture, leveraging a multi-agent swarm to process live audio streams and synthesize complex requirements.

1. The Live Swarm (Real-time Audit)

Powered by Agno (formerly Phidata), the live swarm consists of three specialized agents working in a short-circuit pipeline:

• The Supervisor: Orchestrates the live transcript stream, filtering noise and identifying "Ambiguity Windows."
• The Analyst: Performs deep-reasoning on specific dialogue segments to identify vague quality metrics, undefined references, or scope gaps.
• The Critic: Cross-references Analyst flags against the broader meeting context to ensure high-precision alerts and polite neutralization strategies.

2. The Analytical Boardroom (Post-Meeting)

After a session concludes, the Session Director agent takes over to perform a holistic synthesis:

• Clarity Scoring: Mathematically calculates a grade for the meeting based on the density of resolved vs. unresolved flags.
• Customer DNA Profiling: Reconstructs a psychological and technical profile of the stakeholders (Priorities, Worries, Preferences).
• Digital Replicant Sandbox: Generates a chatable simulation of the meeting persona for follow-up testing.

---

🛠️ Technology Stack

Layer	Technologies
Intelligence	Gemini 1.5 Flash (Live), Gemini 1.5 Pro (Analytical), Agno Swarm Framework
Backend	FastAPI, Uvicorn, Python 3.12, Omium Tracing
Frontend	React, Vite, Tailwind CSS, Framer Motion, Zustand
Integrations	Recall.ai (Video Bridge), Deepgram (Transcription), Firebase (Cloud Registry)
Infrastructure	Railway (Backend), Vercel (Frontend)

---

🚀 Deployment & Environment

Agent Flow is optimized for Zero-Configuration Deployment using a "No-Ngrok" architecture.

Backend (Railway)

The backend autonomously resolves its own callback URLs for Recall.ai webhooks.

Required Variables:

• GEMINI_API_KEY: Google AI Studio key.
• RECALL_API_KEY: Recall.ai API token.
• RECALL_REGION: (e.g., us-west-2) Points to your Recall account region.
• PUBLIC_URL: Set to https://${{RAILWAY_PUBLIC_DOMAIN}}.
• OMIUM_API_KEY: For AI observability and tracing.

Frontend (Vercel)

The frontend connects to the production swarm via environment-aware protocol resolution.

Required Variables:

• VITE_API_BASE_URL: Your Railway production URL.
• VITE_FIREBASE_*: Full set of Firebase configuration keys.

---

💻 Local Development

1. Backend Setup

cd clarity-agent/backend
python -m venv .venv
source .venv/bin/activate  # or .\.venv\Scripts\activate on Windows
pip install -r requirements.txt
uvicorn main:app --reload

2. Frontend Setup

cd clarity-agent/frontend
npm install
npm run dev

---

📊 Key Features & Workflows

Autonomous Join (Calendar Bridge)

While the demo supports manual URL entry, the production architecture is designed to autonomously bridge with Google Calendar, provisioning bots for scheduled syncs without user oversight.

The Ambiguity Matrix

Live alerts are streamed via WebSockets. Each alert contains:

• The Quote: The exact phrase that triggered the flag.
• The Type: (e.g., vague_quality, missing_metric).
• Neutralization Strategy: A polite, actionable question the user can ask right now to fix the ambiguity.

Real-time Agent Tracing

Integrated with Omium, providing a "Trace Hub" where users can see the swarm's internal dialogue and reasoning process in real-time.

---

⚖️ License

Agent Flow is proprietary software built for Biera

Developed by John Pradeep.