coefficient-portfolio

Competitor Connector Outage Intelligence

An agent-based intelligence system that monitors competitor product connectors for outages, issues, and disruptions across multiple sources. Uses AI-powered severity scoring and multi-source validation to identify actionable competitive intelligence.

Architecture

┌────────────────────────────────────────────┐
│ Search Layer (Multi-Source Collection)       │
│ ├── Exa MCP (semantic web/news/forum search)   │
│ ├── Reddit MCP (real-time user complaints)     │
│ ├── Browserbase MCP (marketplace review scraping)│
│ └── Firecrawl MCP (backup web search)          │
└────────────────────────────────────────────┘
                      ↓
┌────────────────────────────────────────────┐
│ Intelligence Layer (AI Analysis)             │
│ ├── Issue Summarizer (1-2 sentence summaries)  │
│ ├── Severity Scorer (HIGH/MEDIUM/LOW + reasoning)│
│ └── Content Filter (exclude docs/tutorials)    │
└────────────────────────────────────────────┘
                      ↓
┌────────────────────────────────────────────┐
│ Data Layer (Storage + Dedup)                  │
│ ├── SQLite (seen issues tracking)              │
│ ├── Date/Content filtering                     │
│ └── Quality validation                         │
└────────────────────────────────────────────┘
                      ↓
┌────────────────────────────────────────────┐
│ Export Layer                                  │
│ └── Google Sheets MCP (structured output)      │
└────────────────────────────────────────────┘

Pipeline Stages

Stage 1: Multi-Source Search

• Exa MCP: Semantic search across web, LinkedIn, forums, news, Twitter
• Reddit MCP: Real-time user complaints from relevant subreddits
• Browserbase MCP: App marketplace review scraping (ratings + text)
• Firecrawl MCP: Backup web/blog coverage search
• Configurable date window (2-60 days back)

Stage 2: Content Filtering

• Exclude documentation, tutorials, release notes, marketing content
• Date range enforcement
• Relevance scoring via Exa semantic matching
• Engagement metrics (upvotes, comments, replies)

Stage 3: AI-Powered Analysis

• Severity scoring: HIGH / MEDIUM / LOW with detailed reasoning
• Issue summarization: 1-2 sentence summaries of each finding
• Error type classification: auth, sync, connection, data_loss, performance
• Status detection: ONGOING / LIKELY_RESOLVED / UNKNOWN
• Business impact assessment: Free-text impact description

Stage 4: Deduplication

• SQLite database tracks seen issues by URL + date hash
• Prevents re-reporting of known issues across runs
• Persistent across multiple scan sessions

Stage 5: Export

• 10-column Google Sheets output
• Sorted by severity (HIGH first) then date (newest first)
• Grouped by connector for easy review

Connector Monitoring

The system monitors 7 competitor connectors across 5 data platforms:

• 2 accounting platform connectors (Google Workspace + Microsoft)
• 1 CRM connector (Google Workspace) [HIGH PRIORITY]
• 1 ERP analytics driver (native)
• 1 marketing connector (Google Workspace) [HIGH PRIORITY]
• 2 data warehouse connectors (Microsoft + Google Workspace)

Each connector has configured:

• Search keywords and patterns
• Marketplace URLs for review scraping
• Relevant subreddits
• Priority level (HIGH/MEDIUM/LOW)

Output Schema (Google Sheets)

Column	Content
Date Found	When the issue was detected
Connector Name	Which competitor connector
Data Source	Platform affected
Issue Description	AI-generated 1-2 sentence summary
Source Type	Web / Support / Reddit / Marketplace
URL	Link to original source
Issue Date	When the issue was reported
Severity	HIGH / MEDIUM / LOW
Status	ONGOING / LIKELY_RESOLVED / UNKNOWN
Notes	Additional context, business impact

Python Module Structure

src/
├── agents/
│   ├── severity_scorer.py    # Claude-powered severity analysis
│   └── issue_summarizer.py   # Claude-powered summarization
├── processors/
│   ├── deduplicator.py       # SQLite-based duplicate tracking
│   └── formatter.py          # Sheets output formatting
├── searchers/
│   ├── exa_searcher.py       # Exa MCP integration
│   └── reddit_searcher.py    # Reddit MCP integration
└── writers/
    └── sheets_writer.py      # Google Sheets MCP output

~1,700 lines of Python across 8 modules.

Results

Test Run (Full 7-Connector Scan):

• Cost: $0.22 total
  • Exa API: $0.17 (9 searches)
  • Browserbase: $0.02 (1 marketplace scrape)
  • Anthropic API: $0.02 (5 AI analyses)
  • Firecrawl: $0.01 (1 search)
  • Reddit + Sheets: $0.00
• False positive rate: 0%
• Content filtering accuracy: 100%
• Date filtering accuracy: 100%
• Successfully identified a major CRM connector outage (48+ hours, 80+ affected instances) with multi-source validation from news, marketplace reviews, and community reports

Projected monthly cost: ~$6.60/month (daily) or ~$0.88/month (weekly)

Stack

• Language: Python 3.11
• AI: Claude API (severity scoring + summarization)
• Search: Exa MCP, Reddit MCP, Firecrawl MCP
• Scraping: Browserbase MCP (marketplace reviews)
• Storage: SQLite (deduplication), Google Sheets MCP (output)
• Config: YAML (connectors + prompts)

Key Decisions

1. Multi-source validation over single-source — Requires corroboration from 2+ sources for HIGH severity. Eliminates false positives from isolated complaints.
2. AI severity scoring over keyword matching — Claude analyzes context to distinguish real outages from routine maintenance. Keyword-only approach had too many false positives.
3. SQLite dedup over in-memory — Persists across runs, so weekly scans don’t re-report known issues.
4. $0.22/scan cost target — Designed for daily or weekly automated runs without budget concerns.
5. YAML-configured connectors — Adding new competitors requires only a config change, not code changes.

Visual Direction

Diagram: Four horizontal layers (Search → Intelligence → Data → Export) with the 4 search sources feeding into the intelligence layer. Show severity classification (HIGH/MEDIUM/LOW) as color-coded badges. Right side panel shows a sample Google Sheets output with connector names and severity indicators. Bottom shows cost breakdown pie chart ($0.22 total).