Private dog context, grounded coaching, and daily care in active development.
Pawly Technologies · applied AI for companion animals
The intelligence layer
for life with a dog.
Pawly combines dog-specific computer vision, grounded behavioral guidance, and live place intelligence in one iOS product. The app is the entry point; the models, longitudinal context, and data infrastructure are the platform beneath it.
Working iOS build
Custom dog-vision models
Deployed location system
Company and product overview · active development
Working systems · measured results · explicit limits
Custom pose and behavior systems with bounded evaluation results.
Live location resolution against named official environmental sources.
Pawly Technologies, Inc. — an independent Delaware software company.
The company thesis
The relationship is the operating system.
Life with a dog is fragmented across behavior, routines, health context, and place. The owner is left to connect it all.
Pawly is building one persistent intelligence layer around the individual dog and the person responsible for them. The consumer app is the entry point; dog-specific sensing, grounded reasoning, and longitudinal context are the reusable infrastructure underneath.
The product surface
One app. Three intelligence systems.
Pawly brings persistent dog context, grounded behavioral guidance, and phone-camera perception into one iOS experience. The interfaces and model output below come from the work in progress.
01
Grounded behavior coach
A private coach that remembers the dog.
The coach combines the dog’s own history with retrieved trainer and research material, then returns a specific action, its reasoning, and the source boundaries.
- Persistent dog-specific context
- Curated retrieval with citations
- Guidance with explicit safety limits
02
Longitudinal context
The relationship becomes structured signal.
Lightweight check-ins, care actions, routines, and questions build useful context before a problem crests. The interface is designed to learn one dog over time without turning care into a score.
One dog, understood over time—not a series of disconnected questions.
Actual model output · outdoors, unstaged
03
See the measured results →
Dog-specific computer vision
General quadruped vision was not good enough.
Pawly’s vision pipeline reads canine pose and movement from phone video. We built it after standard quadruped models produced confident anatomical errors on real dog footage.
Watch the top-right state label
The state changes before the visible movement.
05.6s STRESSED
06.1–06.2s forward burst
In this clip, the model changes state while the dog is still planted—roughly half a second before the visible forward burst.
Deployed infrastructure
The government posts it. Your dog still swims there.
Pawly reads the official record for the water you actually walk, keeps the limits the source itself states, and turns it into one sentence before the next walk. Both incidents below are real, and the read is live.
Pawly Water Watch — live official-source demonstration
Reading the official sources now…
Fishing Vessel Sinking in Snohomish River; Everett, Washington
A 75-foot fishing vessel was reported sinking at a dock in the Snohomish River while carrying 250 gallons of diesel fuel and other oils. State responders, the Coast Guard, and a salvage team were responding.
Open the government recordYou take Lucas to the Snohomish River trail, and he sometimes gets into the water.
What this means for Lucas Keep Lucas out of the Snohomish River near the reported incident for now. A fishing vessel carrying diesel and other oils sank there. This is an incident report, not a public-health closure, so check the official response update before your next river walk.
NOAA IncidentNews covers selected incidents where NOAA provided scientific support; it is not a complete list of U.S. spills. The official read is live (live refresh in progress). The saved place and Lucas’s routine are example owner context (example place + Lucas’s water habit). Delivery to owners is not turned on yet. No match is ever presented as an all-clear.
Deployed now
Place resolution, official-source matching, incident classification, hydrologic context, plain-language action, and direct citation.
Product delivery next
Saved places, background monitoring, deduplication, opt-in circles, and push notifications to the people whose dogs use that place.
The same pattern beyond water
Different evidence. One plan for this dog.
Water Watch is one version of the engine. Pawly can apply the same transformation to food, routines, behavior, and health records: preserve what each source can actually prove, combine it with the dog’s context, and produce a small plan instead of another pile of data.
Inputs Pawly can join
Use food rewards without quietly adding more food on top of his normal day.
Matched in Pawly’s 7,162-product catalog with 54 ingredients traced to the stored label source.
AAFCO definitions, FDA ingredient context, WSAVA nutrition guidance, and Pawly’s cited coaching corpus.
Example composed output
Plan, not diagnosis
Lucas’s food + training plan
-
01
Keep the base food decision unchanged for now.
An ingredient scan can flag questions, but it cannot establish that the whole diet is adequate or deficient.
-
02
Measure his normal daily portion first.
Set aside part of that same food for routine training rewards instead of adding every repetition on top.
-
03
Save separate high-value treats for harder work.
Record them as extras so the next check-in can compare the plan with Lucas’s weight and routine over time.
-
04
Do not add calcium—or any supplement—from this scan alone.
Confirm the package’s nutritional-adequacy statement and life stage. Bring the label and feeding amount to his veterinarian before changing nutrients.
What Pawly refuses to invent
No deficiency claim from an ingredient list. No eggshell recommendation. No dose without the complete diet, life stage, amount fed, and veterinary context.
This is a code-native example of the next composition layer using systems and data already in Pawly’s workspace. The product match and ingredient analysis exist today; longitudinal diet-plan tracking is not yet connected.
Current structured-data inventory
- 7,162dog food products cataloged
- 497graded corpus sources
- 56regions covered — 50 states, DC, territories
- 55regional planning cards
- 59threat mechanisms in the longevity funnel
- 156breeds in the longevity spine
- 603life-table rows across 33 strata (Teng 2022, CC BY 4.0)
- 122parent-club breed health statements extracted
- 40breed sheets shipping in the backend engine
Every figure above is counted from a source file in the current build, not estimated.
Technical diligence
Measured systems. Scoped claims.
Models trained, corpora assembled, and results measured. Every result below names the bounded evaluation or inventory it came from.
Expert-training words
Across 681 professional dog-training transcripts.
Source documents
497 graded corpus sources plus 197 companion-wellness documents.
Reviewed posture images
Hand-validated before model training.
Public discussions
Collected across 89 communities, 2008–2026; 1,376,863 screened and 283,064 read and structured in depth. See how the set was narrowed.
Measured computer vision
Computer-vision evaluation.
Model
Result
Scope
Dog-pose estimation
0.990 / 0.858
Detection / keypoint mAP@50, custom 26-keypoint model, 100 epochs.
Behavior-state read
86.4%
Held-out accuracy on the measured four-class evaluation.
Vocalization detection
AST
Detection built on AudioSet-trained AST; no bark-emotion accuracy claim.
Measured AI coach
Grounded-answer evaluation.
Metric
Result
Scope
Retrieval corpus
10,675
Curated trainer and research chunks in the production retrieval corpus.
Answer accuracy
4.88 / 5
Adversarial evaluation across behavior, training, grounding, and safety.
Grounded in source
92.5%
Answers traceable to retrieved passages in the named evaluation.
Hallucination / safety
0% / 100%
Invented facts / rule adherence on that same bounded test set.
Benchmark depth
384
Graded responses across an 11-model adversarial benchmark.
Why build a dog-specific model? The standard academic quadruped model confidently fired deer-specific anatomy on real dog clips. Confidently wrong is worse than uncertain, so we trained our own.
Read the research and limitations →The platform strategy
Product wedge and expansion path.
Persistent context for one dog
Behavior, routines, questions, progress, and place become one private longitudinal record rather than disconnected transactions.
Multimodal understanding
Phone video, behavior questions, routines, and supported audio detection are designed to meet in one dog-specific context layer.
Location intelligence, then delivery
Official-source resolution works now. Saved-place monitoring, deduplication, opt-in circles, and push delivery are the next product layer.
Market understanding
The roadmap starts with observed owner behavior.
Pawly’s product decisions are informed by large-scale aggregate analysis of public dog-owner discussions and millions of words of professional training material. The methodology, findings, effect sizes, and limitations are published for independent review.
Company
Pawly Technologies, Inc.
Pawly Technologies, Inc. is an independent Delaware software company building consumer AI products. The company is developing Pawly and the dog-intelligence systems beneath it.