All work

Case study · 04 / 04

AI Automation & n8n

Production AI workflows: code review pipelines, content generation, custom MCP tooling — used in real operations, not as demos.

Role
Engineer + workflow designer
Timeline
Ongoing
Status
In production

The problem

Most teams want “AI” in their workflows but end up with chatbot widgets that don’t connect to anything operational. The interesting work is the opposite: pipelines where an LLM is one step among many, with strict input contracts, deterministic glue, and an auditable trail.

The gap is rarely the model — it’s the orchestration. n8n + Claude + a custom MCP layer covers most real-world cases without standing up a full agent framework.

My approach

n8n as the orchestrator — visible, debuggable, replayable. Claude (or an OpenAI / Groq model where it makes sense) as the reasoning step. Custom MCP servers expose internal systems as typed tools so prompts stay short and grounded in real data instead of pasted JSON dumps.

Every workflow ships with a definition-of-done: which inputs trigger it, which outputs are expected, what happens on failure, who gets paged. Workflows are versioned in git (n8n export), not edited live in a UI.

Delivered

What I shipped.

  1. 01

    AI code-review pipeline

    PR-triggered review that pulls diff context, runs a structured review prompt against Claude, and posts findings as PR comments — with tone and depth tuned for the team’s actual standards, not a one-size review.

  2. 02

    Content generation pipelines

    Listing copy, marketing variants, internal docs — typed templates + locale targeting + post-generation lint pass so output ships at quality, not just at length.

  3. 03

    Custom MCP tooling

    Internal systems exposed as MCP tools so Claude (Desktop and Code) can answer real operational questions against live data — read-only by default, mutating tools opt-in.

  4. 04

    Versioned workflow library

    n8n workflows exported and committed; changes go through PR review like any other infrastructure change.

Decisions

Worth calling out.

  1. 01Why orchestration > agent frameworks (for now)

    Full agent frameworks promise autonomy but cost predictability. n8n + a focused MCP surface gets you most of the value with workflows that are visible, replayable, and debuggable when something goes wrong at 11pm.

  2. 02Workflows belong in git

    Editing workflows live in a UI is fine for prototypes — it stops being fine the moment a production pipeline depends on them. Export, commit, review, deploy — same loop as code.

Outcome

Multiple workflows running in production for code review, content generation, and operational lookups. Continuous expansion as new MCP-shaped surfaces become useful.

Stack

  • n8n
  • Anthropic Claude
  • Claude Desktop
  • Claude Code
  • OpenAI SDK
  • Groq
  • Model Context Protocol
  • TypeScript
  • Node.js

Working on something similar?

Free 15-minute discovery call. We’ll work out whether I’m the right person for it, and if not, I’ll point you somewhere better.

Book a discovery callEmail me

Next case study

Ajar

PropTech platform — a listing marketplace I architected solo, plus engineering across the Ajar property-management suite.

Read the case study
AI Automation & n8n · Alex Thomas