Altaviz — AI Media Buying Copilot

Live: altaviz.vercel.app · MCP: https://altaviz.vercel.app/api/mcp

Built by David Fernandez for the It's Today Media Build Challenge.

60-second tour: open the demo → expand the Google tracking-outage finding → ask the copilot "What should I kill today?" → approve its actions in the queue.

What it does

Watches a multi-platform ad account (Meta, Google, Taboola, TikTok) and closes the gap between when something breaks and when a human notices.

• Detect — a statistical engine (plain TypeScript, no LLM in the loop) finds the five failure modes that cost affiliate teams money: creative fatigue, CPA drift, spend spikes, conversion-tracking outages, underfunded winners. Every finding ships with evidence and a $/day impact.
• Decide — a Claude copilot (streaming tool-use agent, claude-sonnet-5) grounded in the same tools. "What should I kill today?" gets dollars, evidence, and reasoning — it knows a tracking outage means "fix the pixel", not "cut spend".
• Act, approved — the agent proposes typed actions with exact platform-API params. They queue for human approval. Nothing executes without a click.
• MCP — the identical tool registry is a Model Context Protocol server, so the team can drive the account from Claude Desktop, Claude Code, or Cursor.

claude mcp add --transport http altaviz https://altaviz.vercel.app/api/mcp

<details> <summary>Claude Desktop / Cursor config</summary>

{
  "mcpServers": {
    "altaviz": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://altaviz.vercel.app/api/mcp"]
    }
  }
}

</details>

Demo data: a deterministic seeded account — 6 campaigns, 90 days, realistic affiliate economics (CPA at 60–85% of payout), five injected anomalies. No database; every visitor sees the same story. Real platform adapters swap in behind the same Account → Campaign → AdSet → Ad → DailyMetric interface.

Why this tool

Your business buys media to build email/SMS lists; ROI is the metric. The expensive failures at that scale aren't strategy — they're detection lag. Dashboards show numbers, not what changed, what it costs per day, and what to do. I chose this over a creative generator or LP builder deliberately: those make more assets; this protects margin on every dollar already being spent — and it's the class of system I've built for five years (anomaly detection on streaming data, agentic tool-use).

Design decisions:

• Statistics, not prompts. z-scores, trend slopes, significance gates in lib/detect. LLMs narrate and decide; they never invent numbers. Findings are reproducible and auditable.
• List economics built in. Pause/scale calls are priced on payout plus a configurable backend $/lead — list-building campaigns often run at front-end breakeven on purpose, and the detectors know that.
• One tool registry, two consumers. lib/tools defines each tool once (zod schema + executor); the chat agent and MCP server both consume it.
• Hard approval boundary. The agent can propose raise_budget; it cannot call it. Product decision, not a gap.
• Public-demo hardening: per-IP rate limiting, capped agent turns/tokens/history.

What I'd build next

1. Weeks 1–2: read-only Meta/Google Ads API adapters behind the existing interface; validate detector thresholds against incidents your buyers remember.
2. Weeks 3–4: scheduled morning briefing to Slack with approve buttons; wire approved actions to platform APIs behind the same HITL boundary, with an audit log.
3. Month 2: close the creative loop — fatigue finding → refresh brief → your video-gen pipeline → upload as paused drafts via your ads MCP server.
4. Ongoing: eval suites — detector precision/recall on known incidents; copilot recommendations scored against what senior buyers actually did.

Architecture

Next.js 16 · TypeScript · Tailwind · @anthropic-ai/sdk · mcp-handler · zod · React Three Fiber (hero) · Vercel.

lib/data/             seeded account generator + types
lib/detect/           statistical anomaly engine (pure functions)
lib/tools/            ONE zod tool registry → consumed by both ↓
app/api/chat/         streaming Claude agent loop (NDJSON)
app/api/[transport]/  MCP server (Streamable HTTP)
app/app/              dashboard · app/  landing

npm install
ANTHROPIC_API_KEY=sk-ant-... npm run dev   # everything but chat works without a key
npx tsx scripts/sanity.ts                  # prints KPIs + anomaly findings, no server

Known limits (deliberate for a demo): in-memory rate limiting (per serverless instance); chat history replays as text only; MCP is unauthenticated read-only demo data.