Autonomous Competitive Intelligence Agent That Self-Improves

Built for the Self-Improving Agents Hackathon — NYC 2.21.2026

WebWatcher is an autonomous AI agent that:

• Monitors competitor websites in real-time using Google DeepMind's Gemini 3 Pro
• Detects strategic changes (pricing, features, messaging)
• Scores relevance using Braintrust's evaluation framework
• Self-Improves by rewriting its own detection prompts when scores drop
• Visualizes intelligence in a live dashboard with embedded charts

The "Wow" Factor: Watch the agent autonomously rewrite its own prompt file when it detects low relevance scores — true self-improvement in action.

• Google DeepMind — Gemini 2.0 Flash (URL reading, change detection)
• Braintrust — Self-improvement engine (scoring, prompt evolution)
• Lightdash — Live intelligence dashboard (embedded charts)

• Backend: Node.js + TypeScript + Express
• Frontend: React + TypeScript + Vite
• Visualization: Recharts
• Storage: JSON files (lightweight)

# Backend
npm install

# Frontend
cd frontend && npm install && cd ..

cp .env.example .env
# Fill in your API keys in .env

• Gemini: https://aistudio.google.com/apikey
• Braintrust: https://www.braintrust.dev → Settings → API Keys
• Lightdash: https://app.lightdash.cloud → Settings → API tokens
• Airia: https://app.airia.ai → Settings → API Keys

# Terminal 1: Backend API
npm run dev:api

# Terminal 2: Agent (optional, runs on schedule)
npm run dev:agent

# Terminal 3: Frontend
npm run dev:frontend

npm run seed

TopDog-Competitor-WebWatcher/
├── agent/              # Core agent logic
│   ├── index.ts        # Main agent loop
│   ├── gemini.ts       # Gemini integration
│   ├── braintrust.ts   # Braintrust scoring + self-improvement
│   ├── lightdash.ts   # Lightdash integration
│   ├── prompts.ts      # Detection prompts (self-rewritten)
│   └── store.ts        # JSON file storage
├── api/                # Express REST API
│   ├── server.ts       # Express app
│   └── routes/         # API endpoints
├── frontend/           # React dashboard
│   └── src/
│       ├── components/ # UI components
│       └── lib/        # API client
├── data/               # JSON storage (signals, scores, history)
├── scripts/            # Utility scripts (seed, etc.)
└── .env                # API keys (not in git)

1. 0:00-0:30 — Show dashboard with 3 competitor URLs
2. 0:30-1:30 — Click "Run Agent", watch signals appear
3. 1:30-2:30 — Show self-improvement (prompt rewrite)
4. 2:30-3:00 — Show score trends and impact

• ✅ Real-time competitor monitoring
• ✅ Strategic change detection (pricing, features, messaging)
• ✅ Relevance scoring via Braintrust
• ✅ Autonomous prompt rewriting
• ✅ Live dashboard visualization
• ✅ Real-time updates (1-5 second polling)

• Autonomy: ✅ Agent runs without manual intervention, self-improves
• Idea: ✅ Solves real competitive intelligence pain point
• Technical Implementation: ✅ Clean architecture, 3 sponsor tools integrated
• Tool Use: ✅ Google DeepMind + Braintrust + Lightdash
• Presentation: ✅ 3-minute demo with visual wow factor

• See PROJECT_SPECIFICATION.md for detailed specs
• See BUILD_CHECKLIST.md for step-by-step build guide
• Pause after each step to test before proceeding

CORS Error: Add app.use(cors({ origin: '*' })) to api/server.ts

API Keys Not Working: Double-check .env file, ensure no extra spaces

Agent Too Slow: Use seed data for demo, show one live call

Lightdash Not Configured: Use embedded Recharts as fallback

Built for hackathon submission — 2.21.2026

Built with ❤️ for the Self-Improving Agents Hackathon