GOOSE
Dr Andrea Lattarulo — General Practice, New Zealand
Italy, November 2020, second Covid wave.
Nobody wanted to go to the hospital, especially with Covid. The rumor was of ending up dying alone in an isolation room, holding a warm saline-filled glove as the last "human" contact.
So patients called us, their GPs. They called if they had symptoms, if they tested positive, if they'd been in contact, if they were scared. They called.
Seven, eight hours of consecutive phone calls, one after another. No breaks.
Every phone call, ultimately, was a decision to make: sick enough to need to go to the hospital or safe to be managed at home.
Your brain would melt. Literally. Out the ear. Down the collar.
One Thursday evening, 9:30pm, well past exhaustion, backpack already on my shoulder, just trying to go home and eat dinner half-conscious and sleep to get ready for the next day. A few reports needed checking first. I skimmed through them fast, the way you learn to do in these environments.
A chest X-ray, skimmed the way I'd skimmed the ones before it. Something stopped me and made me start again. And then again. The final time forcing myself to read it with my finger on each word, like a first grader.
3cm apical spiculated lesion; a clearly cancerous lesion I had nearly missed. Twice.
Not because I'm a bad doctor, but because I had been burned out by a system that wasn't built to protect the operator.
No patient should pay for a system failure. Nobody would board a plane whose pilot hadn't been given a chance to rest.
What the second wave was like
When the first cases hit China, most of us didn't give it much credit. I honestly didn't either.
Then it reached Italy, Veneto, my wife's region. In a week we went from "yeah, whatever" to seeing dead bodies carried away by the truckloads.
We survived the first wave with good isolation and a lockdown with no precedent. By the time it passed, we had a plan: trace contacts hard, swab hard, isolate hard. Countless swabs, phone calls, contact tracing.
The first wave crushed hospitals. The second wave crushed GPs first, hospitals after.
Four or five hours of "normal" clinic, then phone calls. Countless phone calls.
We started wearing headphones because the phones would get genuinely hot to touch, and you couldn't tell anymore if the headache was the stress, the sleep loss, the caffeine, or all of it at once.
I was personally responsible for sixty, eighty, sometimes over a hundred active positive patients at a time. This is just my small piece of a much larger picture.
Colleagues died. GPs and hospitalists alike — like in war, we were all brothers in arms.
Everyone did what they could. Dermatologists took anaesthesia shifts. Internal medicine covered the ICUs.
The car
One day I arrived at my clinic at 1pm and couldn't find parking, so I left the car with the blinkers on outside someone's house, meaning to move it within the half hour.
I walked out at 9:45pm.
The local police were standing next to my car, called out by the homeowner. I had completely forgotten it existed.
I told them I was the lead GP of the clinic, and the day had swallowed me whole enough that I'd forgotten I even had a car.
They looked at my face — compassion, gratitude, pity, all at once — and told me to go home.
That's a real story, as surreal as it sounds. I still get goosebumps remembering it.
After
After I realized what had happened, my legs actually gave out and I landed hard in the chair. It took a few minutes before I could stand up properly and leave.
Left undiagnosed, it would likely have gone unnoticed for months. It turned out to be stage III. She was still alive and still being treated when I left Italy a year and a half later.
That night stayed with me long after the wave passed.
What Goose Is
Years later, LLMs arrived, and I realized I had the option of building something that could help those near misses not happen.
Being a millennial ex-Navy officer, the Top Gun reference was obvious. I didn't want to design the pilot, I'm the pilot. I wanted to design the guy in the back seat, watching my six, catching what was outside my field of view.
Goose is my clinical wingman. Not the guy flying the plane, I hold that seat, always.
The clinical and medicolegal responsibility rests with me, the physician, always. Period.
Goose is the one in the back, catching what a rushed brain misses at consult twenty-something on a Friday. I've run it daily, on real patients, for over a year, iterating it, refining it, keeping faithful to my original intent.
It raises differentials I hadn't considered. Not because it understands medicine better than I do, most of the time it doesn't, but because it gets there by a different route. Years of seeing patients build one kind of pattern recognition. Training on an enormous spread of cases builds another. Different minds, different paths to the same table.
It stress-tests my reasoning before I act on it, not after. Closer to discussing a case with a colleague than taking an instruction from a system, if I disagree with what it raises, I say so, and I only act once I'm actually convinced.
The call is always mine. Because the medicolegal responsibility is mine, regardless.
It's not that the model doesn't know the medicine. Ask it the same clinical question cold and it would probably still get there. The problem was never a knowledge gap. It's interpreting messy data, weighing possibilities, and making sure nothing gets missed — under exactly the conditions where that's hardest to guarantee.
What Makes Goose Different, in full
Goose runs on general-purpose LLMs. I haven't had the chance to fine-tune a model specifically for this, and I think if I could, it would perform even better.
"So it's just a complex prompt for an LLM?" Yes and no.
Anyone who's worked with LLMs knows context is most of what determines output quality.
Goose's real contribution is context, deliberately shaped: not information the model didn't have, but the why behind how it should reason here.
I called that set of instructions Core Values, on purpose, because that's what they actually are.
There are countless examples I could give — a diagnosis I hadn't considered, a case reasoned through more carefully.
The most common win is quieter. The case that keeps eating at your mind on the drive home, because most doctors genuinely care what happens to their patients.
Being able to log on and review it, much like I would with a colleague, helps confirm I did everything — and if I missed something, it's a chance to patch it before any harm is done.
None of this produces hard recommendations. It's closer to a conversation with a sharp senior colleague than an instruction from a system.
If I disagree with something raised, I argue it, and I only act once I'm actually convinced — because I'm the one ordering the test, making the referral, prescribing the medication.
The limitation
There's a limitation worth naming plainly, maybe the most important one: this works because of the fifteen years of medical experience I bring to it.
I'm honestly not sure how well it would work for someone with less experience, because machine bias is a bigger problem than we realize.
We don't usually turn to AI for answers we already have, but for cases where we're uncertain — exactly where a confident hallucination is most likely to go undetected, and cause real harm.
A discussion is only a discussion between minds of comparable standing. If a junior doctor is "discussing" a case with a senior consultant, that's not really a discussion, it's being lectured, and it's the junior doctor's job to just absorb it.
The same risk sits underneath every session with Goose. It only works as a real second opinion if I'm actually willing to take my own reasoning apart — the same mental openness you'd need at a multidisciplinary team meeting, willing to notice when you've fallen in love with your first suspicion. Every doctor does this. We're trained not to, and only years of practice really teach you to catch yourself doing it anyway.
There's a second limitation underneath the first: how cautious it should actually be.
A system tuned to overcall risks over-testing and overdiagnosis, real harm, just a different kind. A system tuned to undercall risks missing the one thing that mattered. Finding that line is most of what "the art of medicine" actually means, and it's something doctors learn over years, not something I'd claim to have solved by writing a prompt.
Mechanism outline
A few of the disciplines it runs on, at the level that matters, not the level that would let anyone rebuild it:
- A hallucination lock, so nothing gets stated as fact unless it can actually be traced back to the case.
- A confidence-honesty rule, so it tells me when it's genuinely unsure, instead of sounding certain by default.
- A silent omission check, aimed at the small set of dangerous things that don't announce themselves.
- A pushback protocol, so a real disagreement doesn't just quietly get smoothed over.
- Case isolation, so nothing from one patient bleeds into the next.
Outline only, deliberately. What made these actually hold under real cases took more than forty iterated versions, and the mechanism design itself isn't published.
What's harder to replicate than any of this is the year that shaped it.
The 15-Minute Reality
A normal GP consult is about fifteen minutes. You listen to the patient, pull a coherent story out of what's often partial and murky, work out what matters, rule out what you can't afford to miss, examine, decide what tests are needed, weigh medication against history, make the call, and document it precisely enough someone could read it in ten years and see exactly what was done and why.
In about as much time as many people take to leave a mean comment on YouTube while taking a dump.
And you do it over twenty times a day.
It shouldn't come as a surprise that things can slip. Even if you're good.
Why It's Not a Product
I studied the regulations and understood the SAMD problem, and I didn't want to find a workaround for it.
That takes people, time, and money I didn't have alone, and I couldn't convince anyone to invest in doing it right. So I parked the product and kept flying the prototype.
The regulatory story
I didn't want to guess at this, so I did the work properly: the Stanford AI in Healthcare specialisation, a masterclass on the EU AI Act, coursework through Johns Hopkins on healthcare data privacy, trustworthy AI in healthcare through Politecnico di Milano. A couple of months of real study, on top of a long tail of reading since.
Software as a Medical Device isn't defined by what you call your tool. It's defined by what it does.
I pitched a partnership to a company I was working with on their own EMR-integrated AI tool. It went nowhere. I tried building visibility on LinkedIn. I logged everything — versions, changes, cases.
Nothing moved. No doors opened. I had poured everything into it and got silence back.
I stepped away from the pitching for a long stretch. I kept using Goose and improving it daily throughout, because by then I didn't feel like as good a doctor without it.
Closing
A GP who has run a real clinical AI tool, daily, on real patients, through more than forty iterated versions, for over a year.
I'm not an AI developer. I'm learning how to get the most out of these tools for something that actually helps me and my patients.
I want it done right, with good conscience and an ethical mindset, or not done at all.
If you share that vision, I'd be happy to talk, whether it's about Goose or not, as long as it's done properly.
Get in touch