How I used AI to find what two service calls missed

Wed, 29 Jul 2026 14:21:53 -0400

My heat pumps stopped making cold air, and for a year nobody could tell me why. Not the HVAC company I paid $1,150, not Daikin’s support line, and definitely not me.

Here’s the part that still gets me: I don’t know anything about heat pumps. A week ago I couldn’t have told you what an expansion valve was. I called it, out loud, into my phone, “one of the valves that controls if something heats or something cools.”

Three days later I had a stored error code, a month of charted data, and an 11-page document I can hand to a technician. This is how that happened.

What’s broken

We’ve got two Daikin outdoor units from 2020, and each one feeds two indoor heads. The upstairs zones stop making cold air while the fan keeps blowing, and nothing shows up on the display, no beep or code or blinking light. Downstairs, on the same outdoor units, everything cools fine all summer, and heating has never failed once in six Vermont winters.

For years I had a workaround. I built an Apple Shortcut that flipped the upstairs heads into heat mode for 180 seconds and then back to cooling, and it worked until it didn’t. After that, the only reliable fix was walking outside and pulling the disconnect switch, which buys a few hours.

What two service calls missed

In July 2025 they charged me $450 to diagnose a bad expansion valve, and that August two techs replaced the metering hardware on one unit, spent hours in my driveway on hold with Daikin, and billed me $1,235.99. It dropped to $700 only after I pointed out they hadn’t applied the 12-year parts warranty they’d just spent the afternoon confirming.

The part they pulled out, hand-marked UPSTAIRS. I kept it, which turned out to matter.

I asked about error codes. They looked at the indoor unit’s status lamp, saw nothing blinking, and told me there weren’t any. I believed that for a year.

Their parting theory was that Green Mountain Power’s demand response was cycling my equipment. Their notes read, in part: “Ran about 15 minutes and had proper running pressures and proper cooling.”

That’s the whole problem in one sentence, because you can run it for 15 minutes and it works fine, and everything I was actually complaining about happened at hour three.

Talking it through

I opened a session with Claude and just talked. Not a tidy summary, the actual rambling version, with the ums and the wrong words and a complaint about the driveway hold time thrown in.

That messiness turned out to be the point. I mentioned the whooshing sound I used to hear and don’t anymore. I mentioned that the kids' room takes ten minutes to come up to full cooling. I didn’t know either detail mattered, and I couldn’t have known. Both of them ended up in the final report.

Then it read the manual. Daikin publishes 200 pages of service documentation for these units, and on pages 100 through 105 there’s a procedure for pulling stored error codes with the wireless remote. You hold TIMER CANCEL for five seconds and the unit beeps through the code alphabet.

I went upstairs with the remote, and both problem heads gave one long beep at A5, which is freeze-up protection.

So they do store codes, and they’d been storing them the whole time, and nobody ever thought to ask.

Page 140 explained the rest. A single A5 clears itself in three minutes, but “if the error repeats, the system is shut down.” That’s why my heat trick worked for years and then quit on me. Warming the coil clears one event but does nothing for a repeat shutdown, and pulling the disconnect clears the latch and re-homes the valves, which is why that was always the thing that worked.

Turning the house into an instrument

I have Sensibo bridges on all four heads reporting temperature and humidity every 90 seconds, and I’d been using them as a thermostat app.

The single most useful thing I got all week wasn’t a fact, it was a test. A coil cold enough to cool a room is also cold enough to pull water out of the air. If the temperature is flat and the humidity is flat, that coil isn’t doing anything, whatever the fan sounds like.

I’d been judging these machines by whether the air felt cold on my hand, which is exactly how you end up with a service call that ends in a shrug, and once the question became “did the humidity drop,” every ambiguous afternoon in my data turned into a plain yes or no.

Bedroom and living room on the same outdoor unit. The temperatures track. Now look at the humidity: bedroom at 70 to 80 percent, living room at 50 to 55.

From there we ran real experiments, and this is the part I couldn’t have designed myself. Switch on the downstairs heads so each outdoor unit has two zones calling. Both compressors ran, both downstairs rooms cooled, and both upstairs rooms did nothing, which killed my own favorite theory about the system not throttling down far enough.

The whole diagnostic day. Downstairs comes on at 13:50 and drops. Upstairs doesn’t move.

Then I went up with an infrared thermometer. The kids' room coil read 80.0°F in a 76.6°F room. A cooling coil should run 20 to 30 degrees below the air around it, and mine was 3.4 degrees warmer. Twenty feet away on the same outdoor unit, the kitchen coil sat at 49.1°F with condensation beading between the fins.

Overnight. Kids' room in red, cycling normally and then quitting at 1:22 a.m. Bedroom in orange, giving up 28 minutes in.

The number that cracked it came the next morning. We ran my heat trick as a measurement instead of a fix, and the bedroom fed for 27 minutes and quit. The night before, a full power cycle had fed it for 28. Two completely different resets, different weather, different load, and the same time to failure within one minute.

Debris floating around in a refrigerant line doesn’t carry a stopwatch. Something closing that valve on a schedule does.

The morning run. It feeds, it quits at 27 minutes, and then the dewpoint never moves again no matter what I do with the fan.

Building the report

This is the part I’d tell a friend about.

All of that lives in an 11-page document now. It opens with a table of my equipment and serial numbers, then the stored codes and how they were pulled, then every measurement with the time it was taken, then the mechanism with page citations back to Daikin’s own manual. It names what’s already been ruled out and why, so nobody repeats work I’ve paid for twice.

It also dates the failure. I paid for a month of history I didn’t know existed, and it put the start of the decline in the first two weeks of July.

It ends with a numbered sequence for the technician. Read the stored codes on all four heads before touching anything. Ohm the port thermistors against the resistance table on page 164. Swap the stator coils to see whether the fault follows the coil. And the one I care about most: after a restart, watch the bedroom port’s valve position for half an hour, because I can tell you it will feed for 28 minutes and then stop, and that will show you whether something is closing the valve or something is blocking it.

There’s a one-page call sheet too, with three authorized dealers and a screening question I ask before booking: will your tech read the written evidence and pull the stored codes before resetting anything? A yes books the visit.

I’ve also got photos of every part that came out of these machines, matched against the part numbers on the boxes.

Two boxes, not one, which contradicted the invoice.

The coils came off too, both on one harness. That narrows what’s left to suspect on that machine.

None of that is a repair, but it’s a repair that can’t be fumbled, and that’s a thing I didn’t have in August.

Where it got things wrong

I wouldn’t trust this post if I skipped this part.

Early on it decided that cold refrigerant lines plus no cooling meant the air wasn’t reaching the coil, so a dirty blower wheel. That was reasonable and it was wrong, and it took a three-hour test with the fan stepped through every speed to kill it off.

It read the invoice and concluded that one valve had been replaced, but I remembered two, and then it decided from my photos that only the valve bodies were swapped, and then I found a third box and it flipped again. Both corrections came from cardboard I’d kept, not from anything it worked out.

It also wrote “six years of failures” into my report because that’s what my rambling first description implied, and that’s wrong. These cooled fine from 2020 through 2024, and the correction matters, because five good years followed by both machines failing in the same summer points at something cumulative instead of one random broken part.

The daily monitor we set up would have lied to me too. It used fixed temperature thresholds and the overnight failures came in underneath them, so the morning report was going to call it a clean night.

It’s good at reading, arithmetic, and remembering everything, and it’s bad at knowing what it hasn’t been told, so that part stayed my job.

Where it stands

I don’t have an answer yet, but I know exactly what to ask.

The kids' room unit meters fine after a power cycle and then latches into freeze protection when the room gets cold and the load drops off. Its valve isn’t broken and I’m not letting anyone sell me one. The bedroom unit has a valve, coil, and harness all under a year old, and it starves within half an hour of any reset. Nobody has ever touched that unit’s control board, port thermistors, strainers, or whatever’s floating around inside the lines.

So that’s the list, and a third valve isn’t on it.

Here’s how I see it

My house was talking the whole time. It stored a code, kept a sensor record, and would have told anyone who asked which port was starved and when.

The information was never the scarce part. What’s scarce is somebody with the patience to read 200 pages of service documentation, the arithmetic to score 120,000 sensor readings, and the willingness to be wrong four times in three days and say so.

I couldn’t have done that a year ago and I still can’t do it alone. What I can do now is walk into a service call with a written record, a stored code, a chart of the week it started, and one specific test I want run.

That’s a very different conversation than the one I had in my driveway last August.

Written July 29, 2026, in Montpelier. The technician hasn’t been out yet. When they have, I’ll report what they find, including if it turns out I was wrong about all of this.

The Sunshine Trail: A Speculative Brand Campaign for Lawson's Finest Liquids

Wed, 11 Feb 2026 11:39:25 -0400

I spend a lot of my free time thinking about how the brands I love in Vermont could push their digital experiences further. I can’t help it! I see a brand doing something great and I immediately start thinking about what else they could build.

I’ll give you an example.

Lawson’s Finest Liquids is one of Vermont’s most beloved craft breweries, and they’re in the middle of expanding into new markets along the East Coast. Their brand is incredible, their community impact is real, and their commitment to freshness is central to everything they do. But that story is mostly told through a traditional website. I kept thinking: what if you could experience that expansion as a journey?

So I built The Sunshine Trail, an interactive map that connects Lawson’s home in Waitsfield, Vermont to Asheville, North Carolina. 791 miles of mountain communities, outdoor adventures, and cold beer.

The map shows every Lawson’s location I could identify (breweries, retailers, bars, restaurants) alongside hiking trails, river put-ins, community partners, and DC fast chargers along a scenic road trip route through the Blue Ridge Parkway, Skyline Drive, and Vermont Route 100. Every design element pulls directly from Lawson’s brand: their colors, their fonts, their custom SVGs, even the signature wave animation from their website. It should feel like something Lawson’s built themselves.

But the details are where it gets fun. Hover over “Sunshine” in the header and god rays shoot across the page. On HDR displays, the sun pushes brightness beyond standard range. Hover over “Cold Beer” and it starts snowing, a nod to their chain of freshness promise. The Sip of Sunshine logo spins with real momentum physics, and clicking it launches emoji bursts. None of these get in the way if you’re just there for the map. That’s the whole point of them being Easter eggs.

The live impact widget projects real historical data forward, showing dollars raised and solar energy generated ticking up in real time. Community partner popups tell actual impact stories with donate buttons linking directly to each organization. An email capture modal triggers after 30 seconds on the route view, creating a natural top-of-funnel opportunity with genuine value exchange: a curated road trip itinerary.

I built the whole thing with Claude as a coding partner, tested it across 15 screen sizes, and added full accessibility support throughout. On mobile, tilting your phone creates wind that blows the snowflakes. The map expands when you tap a marker so you have room to actually read the popup. Every interaction is designed to teach you how to use the site without having to explain anything.

If you want to explore it, the site is live at thesunshinetrail.com. It’s password protected, so use “linkedin” to get in.

Oliver Ames