Ravi Kumar Jangid Get in touch
AI automation · 2024 - present

20+ multi-agent systems for SEO operations

Content refresh, decay pruning, anomaly detection, news publishing and reporting, wired to live data over MCP.

Fintech ยท Fundraising platform

20+agents running in production
6live data sources wired over MCP
5workflow areas covered end to end

The situation

Running a 1,000-article content engine generates a large amount of recurring, judgement-light work: spotting decaying pages, catching traffic anomalies before they compound, keeping reporting current, refreshing content on a cycle.

The problem

That work is genuinely necessary and genuinely repetitive. Done manually it either consumes the strategist's week or quietly stops getting done. Off-the-shelf tools report on the problem but do not act on it, and they do not know the context of a specific site.

Approach

What I
actually did.

In order, with the reasoning behind each step.

01

Started with the workflows, not the technology

Earlier automation work on n8n, Make, Zapier and AirOps taught me which tasks are worth automating: high frequency, clear inputs, a decision rule that can be written down. Anything requiring taste stayed manual.

02

Gave the agents live data over MCP

Each agent connects through MCP servers to Search Console, GA4, SE Ranking, Bing Webmaster and DataForSEO, so decisions are made against current numbers rather than a stale export.

03

Built one agent per job, composed into workflows

Content refresh, decay pruning, anomaly detection, news publishing and reporting each got a dedicated agent with a narrow brief, rather than one general assistant asked to do everything. Narrow briefs are testable and fail visibly.

04

Kept a human at the decision points

Agents surface, draft and flag; publication and strategy calls stay with me. The value is in compressing the analysis, not in removing the judgement.

Outcome

What I took
from it.

The hard part was never the model. It was being honest about which parts of the job are actually rule-based, and resisting the temptation to automate the parts that only look like they are.

Stack
Claude CodeMCP serversSearch ConsoleGA4SE RankingBing WebmasterDataForSEOn8nMakeZapierAirOps
Context on request

Want the version with names and dashboards?

Employer names, client names and the underlying reporting go out with applications, or on request.