Altaviz
AI Media Buying Copilot with Statistical Anomaly Detection + MCP
Tech Stack
Next.js 16 • React 19 • TypeScript • Anthropic Claude • mcp-handler • Recharts • React Three Fiber • Vercel
Overview
Cross-platform ad monitoring that detects anomalies with real statistics, prices them in dollars, and drafts fixes behind human approval.
Problem
Media buying teams lose money in the gap between when a campaign breaks and when a human notices. Fatigued creative doubles CPA for days; a tracking outage burns spend across platforms; a winning ad set stays budget-capped. At affiliate scale that detection lag is a permanent tax on ROI.
Solution
A statistical engine (not LLM guesses) scans every campaign across Meta, Google, Taboola, and TikTok for creative fatigue, CPA drift, spend spikes, tracking outages, and underfunded winners — each priced in $/day. A Claude tool-calling copilot answers "what should I kill today?" in dollars and drafts typed actions that never execute without human approval. The same tool registry ships as an MCP server for Claude Desktop, Claude Code, and Cursor.
Architecture
Seeded Multi-Platform Account → Detection Engine (z-scores, trend slopes, significance gates) → One zod Tool Registry → Claude Streaming Agent + MCP Server → Human-Approval Action Queue