Model capability isn't the product. The layer around it is.

General-purpose AI keeps getting more capable, and people still don't get proportionally more out of it. A chat box leaves most of the work with the person: noticing the problem, framing it, supplying the context, knowing what the tool can even do, judging the answer, turning it into action, and noticing what happened. Few people carry all of that well, and nobody does at 3 a.m. with a crying puppy.

People don't need a smarter chatbot. They need that work done around their actual task and outcome. For a puppy owner, the task keeps changing as the puppy grows, and the outcome is a dog and a household that are doing well.

The person carrying a puppy's care, whether they're getting ready for pickup or already a few weeks in. They already research, watch and try things; the care is there. What's missing is help that fits this puppy right now and keeps track of what they tried and what happened.

In the iOS beta:

• The record. The coach reads the dog's profile, open problems and their notes, past problems with what was done and how it went, recent care logs, food and health check-ins, and what it has learned about the owner and home. The owner can confirm, correct or delete anything it has learned.
• It asks before it advises. The first time a problem comes up, it states what it already knows and asks the few questions that would change the advice. Anything that can't wait comes first.
• Plans from approved guides. A recurring problem becomes a plan with one current step, composed from editorially approved guide pages that match the problem.
• Check-ins that move the plan. Easier, Same, Harder or Not tonight. Good runs advance the step and hard ones step it back. That logic is plain bookkeeping over what the owner reported, not a model's judgment.
• Everyday care writes to the same record: potty, meals and a feeding plan, walks, food and plant checks, stool and room photo checks, symptom check, limp triage, vet records, training practice and photos.

Two starting points: getting a puppy, before there is a dog to track, and already home, where suggestions follow what the owner is actually working on.

The problem map comes from a large study of public owner discussions. The guides come from a graded library of research, veterinary and working-trainer sources. The food checks match against a product catalog.

That is the starting knowledge, and with enough effort someone else could assemble a library like it. What compounds is the loop below.

Every plan keeps three things apart: what we recommended, what the owner actually did, and what happened. The next answer uses all three, so a plan that failed is handled differently from a plan that never got tried.

Across many owners, the same loop becomes a record no one else has: which step was recommended, for which puppy, in which home, whether it was tried, and what happened over the following weeks. A general chatbot never learns the result, and a content site never sees it. We think that record is a stronger long-term advantage than domain knowledge alone: guides and research can be copied, and what actually happened in real homes can't.

Status: the loop runs in the beta. We're pre-launch with testers, so the record it will build is ahead of us, not behind us.

Owning the outcome loop is what makes our own models worth training. A coach tuned on grounded puppy sources, and over time on what actually helped puppies like this one, is something a general model can't copy.

We already build the models. We fine-tune open-weight models with LoRA and QLoRA and serve them on our own GPUs. We trained our own dog-pose model. Our guide pipeline runs on our own hardware, reading the source library and drafting the guides every plan is built from. The Pawly coach model is next in that line, followed by specialists for nutrition, training and health; until it ships, hosted models on Microsoft Azure carry the conversation. Every model and its status.

• A puppy calendar that looks ahead. The app knows the puppy's age today. Next: what each stage usually brings and what's coming, so owners get ready before it becomes a problem. Typical timing with its conditions, never a guaranteed date.
• The next ten minutes. For the hard night: one clear first move, how long to give it, and when to change course.
• Good moments. Remembering the wins an owner shares, and bringing one back in a hard week.
• Vision and our own camera. The same models that read a training rep watch the home: two bodies, one timeline, outcomes seen instead of reported.
• Clinics, boarding and shelters. One camera over a row of kennels: automatic charting, overnight watch, recovery after surgery, and an owner check-in window the clinic controls. It's also where the strongest data lives: many dogs around the clock, with outcomes the building already records.

A consumer subscription, entered in the puppy window. The puppy keeps changing, so what the owner needs next keeps changing too. That is the reason to come back, and it is a thesis we still have to prove with retention data. The dog's record is what keeps the product useful after the puppy stage.

Pre-launch and bootstrapped, with no outside funding. The app is in TestFlight beta with testers. What this page describes as today exists in the beta; making the main owner journeys dependable end to end is the current work. No paying customers yet, and retention and long-term outcomes haven't been measured.