I work the technical end of support at Help Scout, and increasingly on the AI tools the rest of the team leans on every day. From a rainy apartment in Amsterdam.
Over a decade of this now: Geckoboard from 2014 to 2018, Help Scout since. For most of it the job was the conversations themselves: email deliverability, security questionnaires, our public APIs, the cases that don’t get solved by poking at the UI. These days I spend most of my time building, and take only the trickier conversations. Mostly pattern recognition; the best of them a quiet unpicking of something I haven’t seen before.
The building started in the summer of 2025, and the last six months have been the steep part of the curve. Compass, our sidebar copilot, is the most visible piece: it assembles the ten minutes of context a reply usually starts with, then drafts the reply when asked. There’s a hosted MCP server too, a bridge that lets agents like Claude, Cursor, and ChatGPT work directly against a Help Scout workspace; it fell out of an AI evaluator I was building, and I’ve used it on every AI project since. Smaller internal scripts fill in around them.
All of which is to say: I’m slowly angling toward applied AI engineering. Same problems, same curiosity: just closer to the code.
The through-line is that I like getting to the bottom of things. I’m not a developer by training, so a lot of building Compass and the MCP server meant picking up things I’d never touched before (HMAC signatures, JWT validation, prompt-injection hardening) and shipping them anyway. I learn fastest with something real on the line.
Outside of work I spend a lot of time making music. Mostly dark ambient: drones and pads, the kind of sound that rewards careful listening more than clever playing. Turns out the patience to let a chord evolve for twenty minutes is roughly the same patience required to find the one wrong header in a thousand-line log.
The rigAn Ableton Push, mostly used as a thinking aid.
The deskInbox open, notebook open, headphones on, the same glass of water all morning.
The viewAmsterdam, the right building, the wrong side of it.
Reading is the steadiest habit I have: AI papers and the inbox during the day; Jonathan Strange & Mr Norrell and The Infinity Machine in the slower evenings. Two very different shelves, both worth the long sit.
Selected, not complete. Most support work doesn’t belong on a website. These are the ones I still think about.
Claude in the Help Scout sidebar, with permission to call the APIs it needs: the knowledge base, past tickets, customer history, saved replies. It assembles the ten minutes of context a reply usually starts with, and drafts the reply when asked.
Read the whole thingA daily check on whether Help Scout’s AI Answers actually answered. Reads every thread, tags the customer’s real signal, sends samples to Claude for scoring. A 1,100-conversation analysis split “Contact helped” into three honest buckets and fed the team’s improvement queue.
Read the whole thingA hosted MCP server that connects Claude, ChatGPT, and Cursor to a Help Scout workspace. Around fifty tools across the full API: conversations, replies, tags, the knowledge base, reports. I built it, and I maintain it.
Read the whole thingVendor questionnaires, pen-test follow-ups, vulnerability reports. They land on my desk first; I handle most of them and pass the ones that need real security depth to the team, with enough context that they’re not starting cold.
Read the whole thingThe daily work: bounce diagnostics, SPF/DKIM/DMARC alignment, tracing why one recipient’s server decided today was the day. And the documentation that keeps the same question from being asked three times.
Read the whole thingWrite if you work on applied AI, support engineering, or anything in the unglamorous corners of email. If you email, I promise I’ll reply: usually within a day or two (sometimes faster if it’s a deliverability puzzle).
Or find me on LinkedIn ↗ · GitHub ↗