How and why I built Hearth Consult (demo on the front page)
Most of my career it's been just me in the building. If you've ever worked somewhere with 3 or 4 doctors you know the move.. something doesn't add up, you lean into the next room and say "hey look at this one for a second". Half the time they don't even solve it, just saying it out loud to somebody who gets it is enough.
When you're solo there's nobody to lean over to. You figure it out yourself while the next room fills up and a tech is standing in the doorway asking about the dental.
I'd been using Claude for this in a plain chat window for a while. It helped, but it didn't know my formulary, didn't know the patient, and I was retyping the whole history every time. So last week I just asked Claude to build me the version I actually wanted. The first working version took a few hours.
How it works, the short version:
- I type the case like I'd say it. dog, 75 lb, vomiting overnight, sore belly. The first words come back in about a second - it leads with the bottom line, then a ranked list of what it could be, then the one question that would change the plan. That order was on purpose, I wanted what a good colleague does, not a textbook chapter - it searches a pile of free guidelines I gave it (ACVIM and ISCAID consensus statements, RECOVER, IRIS, AAHA) and shows which ones it used - a second, bigger model reads the answer and adds "one more thing to consider".. the thing you'd forget at 5:45 on a Friday
The part I was most careful about was the drug math. AI is good at reasoning through a case and bad at arithmetic, and it'll give you a wrong number with total confidence. So the doses don't come from its memory. They come from my own clinic formulary (the same 217 drugs behind my drug calculator), worked out in mL for that patient, and then plain code checks the math again. If the words and the numbers disagree, the dose gets struck through on screen. Cats get hard stops on the drugs that hurt cats. I still read every one.
The part I'm proudest of.. type @ and the pet's name and it pulls the chart right out of our practice software, today's exam and x-rays included. The owner's name and phone number get stripped out before anything goes to the AI. After that the answer stops being generic, it's working off that patient.
Then the stuff I added because I kept wanting it:
- it reads the short version out loud in my George voice - one click turns the plan into an owner handout - if I tell it we're out of a drug, it plans around that and remembers - ER tools with no AI in them at all.. CPR timer with a metronome, drip and fluid math, tox checks for dogs and cats. On a bad day I don't want a language model anywhere near a CPR count - it works on my phone, which is honestly where I use it most
It runs on a little server at the clinic and it's just for me, so there's nothing to log into. The 2 and a half minute demo is on the front page: https://futureproof.vet/#demo-hearth-consult
I'm going to keep doing these, a quick demo video for each thing I build, so there'll be more.
If you work solo too.. what do you do when a case doesn't add up? And if anyone wants to know how any piece of this works, just ask.
