I build the tooling that stops a support queue from becoming a headcount problem.

Prototype → production

n8n to Python

A customer chatbot, from prototype to production service

The problem

A large share of inbound volume was repeat questions with settled, documented answers — but every one still consumed an agent.

What I built

Prototyped the entire conversation flow in n8n first, so the logic could be argued about and corrected in days rather than sprints. Once it was right, I rewrote it as a standalone Python service for production, and now maintain and extend it with Claude Code and Codex.

Why it matters

The expensive part of an AI feature is working out what it should do. Prototyping in n8n made that part cheap. Only the proven design earned real code.

Retrieval

RAG over
company data

Internal knowledge retrieval for support

The problem

The answers existed — in the knowledge base, in old tickets, in documentation — but finding them took longer than the resolution itself, and first line often didn't know they existed at all.

What I built

A retrieval layer over the company's own documentation and support history, so the right prior answer surfaces when an agent needs it rather than after escalation.

Why it matters

Grounded in internal data, not a general model guessing. It now answers around 80% of the questions raised — and points to the known solution behind each answer instead of inventing one.

In progress

Zendesk copilot,
built in-house

Agent assist without the per-seat licence

The problem

Off-the-shelf copilots are priced per seat, trained on generic data, and can't see the internal systems that actually determine the answer.

What I built

A copilot built in-house on n8n and the Zendesk API — drafting replies from real internal context and summarising long ticket threads, inside the tool agents already work in. No migration, no new interface to learn.

Why it matters

It knows your systems because it was built against them. Working in test now, being hardened for production.

Automation

3 review channels,
1 queue

App Store, Google Play and Trustpilot into Zendesk

The problem

Public reviews across three platforms sat outside the support workflow entirely. Nobody owned them, response times were measured in weeks, and the complaints inside them never reached the people fixing the bugs.

What I built

Six n8n workflows, all in production: App Store, Google Play and Trustpilot reviews pulled into Zendesk as tickets; auto-drafted replies for five-star reviews across all three sources; and prepared App Store reviews published back without a human touch.

Why it matters

Public reviews became a tracked, measurable queue instead of something someone remembered on Fridays — and the positive ones answer themselves.

Observability

Ops reporting,
from scratch

Monitoring, alerting and the reporting layer beneath it

The problem

There was no support framework and no knowledge base — and no shared view of whether the systems were healthy.

What I built

Grafana dashboards and alerting for system performance, plus ownership of the operations department's Metabase reporting. Built the technical knowledge base from nothing and trained first line on troubleshooting and workarounds.

The result

Routine requests automated and workflow efficiency up 20% — with a first line that now resolves more before it ever reaches me.