Built around the owner's record, not a chat box.

What happens is saved as dated observations about the dog: meals, walks, potty trips, check-ins, scan results. Problems are tracked as open or resolved, with what was done and how it went. What the app learns about the owner and home is kept as separate facts the owner can confirm, correct or delete, and confirmed or corrected facts rank first. The coach reads all of this through read-only tools scoped to the active dog, so two dogs' histories don't mix, and a small model condenses long conversations so history stays usable.

The first time a problem comes up, the coach states what it already knows from the record and from what the owner just said, names what it doesn't know that would change the advice, and asks two or three short questions. No plan until they're answered. The exception is anything that can't wait, such as danger, pain or a safety risk, which it says first. Once it knows the dog and the problem, later turns get a straight, short answer.

A plan is composed from editorially approved guide pages that fit the problem, and it has one current step. Each check-in (Easier, Same, Harder, Not tonight) is recorded against that step. Whether the plan advances or steps back is plain bookkeeping over the owner's own reports, not a model's call. A model reads a free-text note only to catch a real physical sign, such as pain or an injury, that the button would hide, and that always counts as a miss.

The recommendation, what the owner actually did and what happened are kept as separate facts. That is the outcome loop the company is built on.

• Health: answers come from retrieved veterinary sources, must validate against a fixed format, and may only cite sources that were actually retrieved. A safety floor sets the minimum urgency; the model can raise it but never lower it. Limp triage is a fixed set of questions with three outcomes: urgent, same day, or watch at home.
• Food: labels are read on the iPhone itself and matched to our product catalogue. Ingredient scoring is rule-based, and barcode lookup has no model in the loop.
• Photos: stool reads sit behind a deterministic colour check, and every photo result says what a photo can and can't show.
• Routing: before anything answers, a router works out what kind of question it is (training, health, food, or something urgent) and sends it down the right path.

• Fine-tuned open-weight models Trained We fine-tune open-weight models with LoRA and QLoRA on our own GPU and serve them ourselves.
• Dog-pose model Trained Trained in-house on hand-checked dog images, after the standard academic animal-pose model labeled real dogs with deer anatomy. It runs in our video pipeline; the pose check itself is a developer tool, not an owner feature.
• Two-body tracking Trained Dog and handler in the same frame, from our pose model plus a human body model.
• Cue grader Trained Lines up spoken cues with posture changes to time each rep.
• Gaze and check-ins Trained What the dog is looking at, and when it looks back at you.
• Guide pipeline Live On our own hardware: embed, retrieve, rerank and judge across the trainer, veterinary and research library to draft guide pages, which are approved before any plan can use them.
• Food-photo vision model Trained A small vision-language model we fine-tuned on food photos.
• On-iPhone reads Live Food-label text is read on the phone with Apple's on-device vision.
• Conversation coach Live Hosted models on Microsoft Azure, working through our record tools and rules.
• Pawly camera Planned Our own indoor camera and firmware, syncing to the owner's phone without a cloud.
• Kennel-row mode Planned One camera over a whole room, a separate read for every dog, and events logged to each dog's chart.
• Pawly coach model and specialists Planned Our own fine-tuned coach plus nutrition, training and health specialists, trained on our grounded sources and, as it builds up, on the outcome record. Local serving of a coach-class model is already wired to a test backend.

Live: running in the product or pipeline today. Trained: our weights exist. Planned: next in line.

Hardware we own: an NVIDIA RTX laptop GPU for training and local serving, and an Intel Arc machine that runs a large open-weight model for the guide pipeline.

Text about dogs is everywhere. What doesn't exist anywhere is a record of what was recommended for a specific puppy, whether the owner tried it, and what happened. As that record grows, it becomes training and evaluation data for our own models: not just what a good answer sounds like, but which answers actually helped. That is where this goes next.

Pose, two-body tracking, cue grading and gaze are developer-lab outputs today, not owner or clinic features yet, and the Pawly camera and Kennel-row mode are planned. Pose outputs are not diagnoses. Health answers are guidance, not a diagnosis, and they say when to call a vet. The app is in TestFlight beta, and not every path has been proven end to end on a phone.