A 4-agent enterprise pipeline that replaces 10+ hours of manual B2B sales research with a single company name.
Sales teams spend 2–3 hours per prospect researching companies, finding decision-makers, analyzing pain points, and crafting personalized outreach. This system does all of it — automatically, consistently, at scale.
Built as a capstone for the Google AI Agents Intensive (Nov 2025), delivering 6 out of 3 required features — all 3 core + 3 bonus.
| Interface | URL |
|---|---|
| 🌐 Production Dashboard | sales-intelligence-agent.vercel.app |
| 🗺️ Agent Orchestration View | /dashboard/workflow |
| 📋 Terminal Log Viewer | /dashboard/logs |
| 📊 Intelligence Overview | /dashboard |
╔══════════════════════════════════════════════════════════════╗
║ USER INPUT ║
║ "Company Name: Acme Corp" ║
╚═══════════════════════════╦══════════════════════════════════╝
║
▼
╔══════════════════════════════════════════════════════════════╗
║ ORCHESTRATOR AGENT ║
║ Coordinates all agents │ Manages memory & state ║
║ Handles errors │ Structured logging ║
╚═══════════════════════════╦══════════════════════════════════╝
║
╔═════════════════╩══════════════════╗
║ SEQUENTIAL AGENT FLOW ║
╚═════════════════╦══════════════════╝
║
┌──────────────────▼──────────────────┐
│ 1. RESEARCH AGENT │
│ Tools: Google Search API │
│ Output: Company overview + news │
└──────────────────┬──────────────────┘
│
┌──────────────────▼──────────────────┐
│ 2. ANALYSIS AGENT │
│ Tools: Gemini 2.0 Flash │
│ Output: Challenges + opportunities │
└──────────────────┬──────────────────┘
│
┌──────────────────▼──────────────────┐
│ 3. CONTACT AGENT │
│ Tools: Search + custom logic │
│ Output: Prioritized contact list │
└──────────────────┬──────────────────┘
│
┌──────────────────▼──────────────────┐
│ 4. OUTREACH AGENT │
│ Tools: Gemini 2.0 Flash │
│ Output: Personalized emails │
└──────────────────┬──────────────────┘
│
┌──────────────────▼──────────────────┐
│ FINAL OUTPUT │
│ 📊 Complete Intelligence Report │
│ 👥 Decision maker contacts │
│ 📧 Ready-to-send email drafts │
└─────────────────────────────────────┘
┌─────────────────────────────────────────────────────────┐
│ CROSS-CUTTING LAYER │
├──────────────┬──────────────┬──────────────┬───────────┤
│ Memory Bank │Session State │Observability │ Error │
│ JSON cache │ Context across│ Latency + │ Handling │
│ across runs │ executions │ token logging│ Recovery │
└──────────────┴──────────────┴──────────────┴───────────┘
| # | Feature | Status | Details |
|---|---|---|---|
| 1 | Multi-Agent System | ✅ Core | 4 specialized agents with class-based abstractions |
| 2 | Tools Integration | ✅ Core | Google Search API + custom business logic tools |
| 3 | Long-Term Memory | ✅ Core | JSON memory bank persists research across sessions |
| 4 | Session & State Management | ✅ Bonus | SessionState maintains context across agent executions |
| 5 | Observability | ✅ Bonus | Per-agent logging: latency, error level, token usage |
| 6 | Gemini Integration | ✅ Bonus | Analysis + Outreach agents powered by Gemini 2.0 Flash |
Result: 6/3 required features — all 3 core + 3 bonus
Built with Next.js 15 App Router, React Flow, and Framer Motion — deployed to Vercel.
- 4 live metrics cards: companies researched, contacts found, emails generated, cache hit rate
- Agent execution timeline with animated status chips per agent
- Company intelligence summary panel
- Memory bank visualization with cached entries + timestamps
- Outreach email preview with per-contact tabs
- Full React Flow agent graph with animated data-flow edges
- Per-node status indicators:
idle → running → complete → error - Animated bezier connections showing pipeline execution state
- Custom node components matching the agent console design system
- Dark monospace terminal UI
- Columns: timestamp │ agent │ level │ message
- Auto-scroll during live execution
- Python 3.8+
- Node.js 18+ (for frontend)
- Google AI Studio API key → Get one here
# 1. Clone
git clone https://github.com/Harshithk951/sales-intelligence-agent
cd sales-intelligence-agent
# 2. Python environment
python3 -m venv venv
source venv/bin/activate # Mac/Linux
# venv\Scripts\activate # Windows
# 3. Install dependencies
pip install -r requirements.txt
# 4. Configure
cp .env.example .env
# Add your GOOGLE_API_KEY to .env
# 5. Run
python main.pycd frontend
npm install
cp .env.example .env.local
# Set NEXT_PUBLIC_API_URL=http://localhost:8000
npm run dev
# → http://localhost:3000reports/ ← JSON intelligence reports per company
logs/ ← agent_YYYYMMDD.log with per-agent metrics
memory_bank.json ← Research cache (persists across runs)
sales-intelligence-agent/
│
├── agents/ # Core agent implementations
│ ├── research_agent.py # Google Search → company overview
│ ├── analysis_agent.py # Gemini → challenges & opportunities
│ ├── contact_agent.py # Search + logic → decision makers
│ └── outreach_agent.py # Gemini → personalized emails
│
├── tools/
│ └── search_tool.py # Google Search API wrapper
│
├── utils/
│ ├── memory.py # Memory bank + session state
│ └── logger.py # Per-agent observability logging
│
├── frontend/ # Next.js 15 production dashboard
│ ├── src/app/
│ │ ├── page.tsx # Landing page
│ │ └── dashboard/ # Dashboard + workflow + logs
│ └── src/components/
│ └── workflow/ # React Flow agent graph
│
├── main.py # Orchestrator entry point
├── requirements.txt
└── .env.example
# .env
# Required
GOOGLE_API_KEY=your_api_key_here
# Optional
MODEL_NAME=gemini-2.0-flash-exp # or gemini-pro
MAX_RETRIES=3
TIMEOUT_SECONDS=30Change number of contacts generated:
# agents/outreach_agent.py, line 45
for contact in contacts[:3]: # ← adjust here📊 SALES INTELLIGENCE REPORT: Acme Corporation
════════════════════════════════════════════════════
🏢 Company: Acme Corporation | Technology/SaaS | 100-500 employees
🎯 Challenges Identified (3):
1. Scaling infrastructure under rapid growth
2. Managing technical debt accumulation
3. Integrating disparate internal systems
👥 Decision Makers Found (3):
• John Smith — Chief Technology Officer
• Sarah Johnson — VP of Engineering
• Michael Chen — Director of Product
📧 Outreach Emails Generated: 3
📁 Report saved → reports/acme_corporation_20251115.json
⚡ Total execution time: 47.3s
💾 Research cached → memory_bank.json
# Test with known companies
python main.py
# Enter: Salesforce
# Verify memory caching (2nd run should be instant)
python main.py
# Enter: Salesforce
# → "Found cached research for Salesforce (saved 38.2s)"| Decision | Choice | Reason |
|---|---|---|
| Agent topology | Sequential | Each agent requires previous agent's output as context |
| Memory layer | JSON file | Portable, zero-dependency, sufficient for demo scale |
| LLM | Gemini 2.0 Flash | Fast, cost-effective, strong at analysis + generation |
| Frontend | Next.js 15 App Router | Production deployment, React Flow support, Vercel-native |
| Observability | File + console logging | Dual output — human-readable in terminal, parseable for dashboard |
| API layer | Stubbed with hot-swap interface | Decouples frontend from backend; single env var to connect real Python backend |
- Wire live Python backend to Next.js dashboard via FastAPI
- Real Google Custom Search API integration (replace mock)
- LinkedIn API for verified contact data
- Async parallel agent execution for sub-10s pipeline
- Email sending via SendGrid/Gmail API
- Multi-company batch processing mode
- Deploy backend to Google Cloud Run
- Vector store integration (ChromaDB) for semantic memory
Harshith Kumar Mannepalli
- Google AI & Kaggle — AI Agents Intensive Course (Nov 2025)
- Google ADK team — Multi-agent orchestration framework
⭐ Star this repo if it helped you build something.