Building & Breaking
I Gave an LLM My Inbox
On building flimap — a Go MCP server that gives an AI agent read/write access to a Protonmail Bridge inbox over IMAP and SMTP.
BEN WELLER
Twenty-six years of shipping software. Currently available. Bring me the thing that’s actually broken.
- location
- Austin, TX
- ben@b7r.dev
- status
- taking freelance engineering work — see below
- philosophy
- small, composable tools, trust boundaries, software that works when the internet doesn't
- tech
- GoNodepgvectorPostgresRedisClaudeGLMKimiMCPOpenAIbashVimzshand whatever the problem demands
Here’s the qualifying round, self-administered:
- Something specific is broken and everyone already knows it — you don’t need a two-week audit to find it, you need someone to just fix it.
- You want it built by the one person writing the code, not a team that hands you a plan and disappears.
- The internal tool, the integration, the automation nobody’s had time to write — that’s the job. Not a strategy deck.
- Nothing longer than the project needs. Direct, not routed through anybody — that’s what lets a small operation run like it isn’t.
If that’s you: ben@b7r.dev. Let’s meet. Tell me what’s broken — I’ll know before we’re done talking whether I’m actually the right fix, not just a warm body with a keyboard.
You need a genuinely good nerd, and you don’t have one on staff. I’ve been that person for 26 years, across industries with nothing in common except the same instinct — find what’s broken, build the fix — pressure-tested from a foam factory floor to a VC-backed coupon startup. “AI-native” wasn’t a category for most of it.
FutureFoam’s Washabelle brand ran its storefront on something I built from scratch, wired into the UNIX boxes running the foam factory floor — decades-old infrastructure, still load-bearing, talking to a modern stack because it had to. I Love Blue Sea needed e-commerce that understood dock-side fulfillment; I built that, and was still there when the company exited. EMS Home Services runs on systems I built end to end — back office, technician’s field app, PBX call routing that knows which work order is calling before anyone picks up.
At Rewind.me — originally a personal history tracker — the fundable idea became a deals platform: surfacing coupons based on AmEx spending and Foursquare check-ins, a genuinely strange, of-its-moment data problem. I built the targeting logic as a full rule DSL, more engineering than the feature strictly needed. It meant the people who understood the business could adjust the rules themselves instead of waiting on me to ship.
I’ll make that trade every time: a little overbuilt, in exchange for the business staying in the hands of people who understand it.
I was at the next desk over from a team building what’s now a well-known financial-connections platform, trying to get it back into fundable shape. That one didn’t raise. Not every bet pays off, but proximity teaches you what a real investor asks for.
Right now, on-site with one client, I’m doing the older-fashioned version of the same job: absorbing the operational load a small team would otherwise hire for.
AI has changed the math on capacity: I’ve got a department of mid-level engineers on call, day or night. What that doesn’t replace is the part that was always the real job — taking a business problem and turning it into the right technical solution. That’s still one brain. Same one that was doing it in a foam factory and a coupon startup, years before “agent” meant anything besides someone who returns your calls.
I write here too, whenever I can find the time — a scribble on the nightstand, a paragraph typed in-flight. AI sometimes helps get the words down.
Building & Breaking
Which Pipe to Bang On
On the old pipe-banging story, an agent's hardening sweep, and the one pipe in the system nobody could hit but me.
Thinking Machines
Horse Questions
On horse questions, tractor questions, and the real skill of telling an AI no.