Independent software studio · Virginia Beach
Software that remembers, and gets it right.
Azalea Intelligence builds consumer apps around records that fill themselves in from everyday conversation. We also publish our research on keeping those records accurate when the sources disagree.
Mochi · weight
Example record-
superseded5.4 kgVet record, PDF
-
current11.2 lbOwner, in chat
-
error112 lbScanned invoice, decimal point lost
-
superseded5.4 kgThe August vet record, imported again
A keep-the-newest rule would have saved 5.4 kg, because that old record arrived last.
Our first app · iPhone, Android, and web
Azalea Pets
An AI companion that gets to know your pet. Talk to it the way you'd text a friend, and it keeps the record: what they eat, what they weigh, what the vet said, and the photos in between.
- A profile and timeline that build themselves from chat and photos
- Medications, vaccines, vet visits, and weight in one place, with reminders
- A summary of the record you can send to your vet in one tap
- Help judging how urgent a symptom is, with emergencies sent straight to a vet
Free to start. Azalea isn't a veterinarian and doesn't diagnose.
Machine learning research · recbench
When two records disagree, which one is true?
A pet's record comes from many places: the owner in chat, a vet's PDF, a scanned invoice, an old file imported late. They contradict each other. Most software keeps whatever arrived last and is fully confident about it.
recbench is our open machine learning benchmark for this problem. It generates synthetic histories with a hidden true timeline and eleven kinds of conflict, from unit mix-ups and scanning errors to stale re-imports. A method is scored on whether it kept the right value and whether it knew when it wasn't sure.
Predictions for each round are written down in the public repository before the runs.
DOI 10.5281/zenodo.23205119
| Method | Right value kept | Confident and wrong | Bad entries caught |
|---|---|---|---|
| Keep the newestthe usual rule | 92.2% | 7.8% | 2% |
| Language modelDeepSeek, one call per record | 93.6% | 4.3% | 80% |
| Feature modelgradient-boosted trees, hand-built features | 94.2% | 1.8% | 83% |
| Our reconciler, v00.62M parameters, trained only on synthetic histories | 93.8% | 4.8% | 85% |
Confident and wrong is the share of answers given at 0.9 confidence or higher that were wrong. Bad entries caught is recall on entries labeled erroneous. Best in each column is underlined. Figures are from RESULTS.md in the repository, version 0.3.
Our small model matches the strongest baselines on accuracy and catches the most bad entries, but it is still too sure of itself when it is wrong. Calibration is the current work, and a paper is in preparation.
Safety rules that survive a model swap
The health rules in Azalea Pets are written once and tested against adversarial prompts every time they change. We run the same tests on every model we use, so a rule that holds on one holds on the next.
The studio
A small studio, on purpose.
Azalea Intelligence is a product studio that also does applied machine learning research. We keep the whole craft in one place: how the models behave, the product design, and the infrastructure underneath. Small surface, high standard. More apps are in development, and they will show up here first.
- Company
- Azalea Intelligence LLC
- Founded
- August 2026
- Based in
- Virginia Beach, Virginia
- Contact
- hello@azaleapets.com
- Records that fill themselves in
- Profiles, timelines, and histories come together from ordinary conversation and photos. Nobody fills out a form.
- Measured before it ships
- We test how often the software is wrong and how sure it was at the time. Being confidently wrong is the failure we care most about.
- Restraint is a feature
- A good companion knows when to ask, when to act, and when to say it doesn't know.