Honest guide to how the ReptiDex AI morph identifier works

New keepers often struggle to identify the visible traits in the animal they just bought, and a lot of the existing tools overpromise. The goal here is not to replace breeder judgment. It is to give breeders and keepers a faster first pass that is honest about uncertainty.

A big part of the point is education. For people who are new to reptiles, learning morphs takes time, repetition, and a lot of exposure to real animals. The identifier is meant to be a fun way to practice pattern recognition, compare what the model sees with what experienced keepers see, and get better at understanding trait expression over time.

It is not a replacement for the knowledge held by breeders and keepers in the community. That knowledge is still the standard. ReptiDex is trying to provide a useful sanity check while someone is learning, or while they are waiting on input from other breeders whose experience should carry more weight than a model output.

If the tool helps a newer keeper ask better questions, notice the right visual details, and learn faster without becoming overconfident, then it is doing the job it was built to do.

A lot of people hear AI and assume it always means massive waste, huge data centers, and careless energy use. That concern is fair. The truth is more specific: different AI systems have very different footprints, and a narrow image-classification tool like this is not the same thing as running a giant general-purpose chatbot for every task.

The environmental cost comes from compute usage. That same underlying compute is also what powers many of the things people already rely on every day, including cloud storage, routine database operations, analytics processing, image handling, and ordinary application logic. AI inference and training sit on that same spectrum. The important question is how much compute a tool uses and how often it uses it.

At ReptiDex, I care deeply about the environment and I would not knowingly ship something that I believed was causing avoidable harm. The identifier is scoped tightly, called only when a user asks for it, and built to do one specific job rather than burn compute constantly in the background.

We also retrain the model sparingly rather than constantly. Training is the more compute-heavy part of a system like this, so being deliberate about when retraining happens matters. That keeps the identifier far closer to a lightweight, targeted tool than a runaway always-on compute sink.

That does not mean the impact is zero, and I do not want to pretend otherwise. It means the right question is whether the tool is useful enough, honest enough, and efficient enough to justify its footprint. That is the standard I hold it to, and if I ever think it is failing that standard, I would rather scale it back than hand-wave the concern away.

ReptiDex uses a vision model trained on labeled reptile photos. When you upload a photo, it returns its best guesses about which visible traits may be present, along with a confidence score for each guess.

It does not guarantee accuracy.

It does not replace breeder judgment.

It does not identify traits that are not visually expressed. Hets cannot be identified from photos.

It does not match animals to owners or pedigrees.

It does not store images for surveillance or facial recognition purposes.

It does not share images without explicit opt-in.

The training data comes from Dusty’s own breeding records and partner contributions that were provided for this purpose. Labels are trait-level breeder labels, not vague marketing names. Retraining happens as the labeled dataset improves. A formal public evaluation write-up is still in progress, so I am not going to pretend the model has precision numbers that I cannot back up yet.

AI features can be turned off at any time inside the app settings.

You can enable or revoke them at any time in settings.

Identification images are retained in the identification system for 30 days, then deleted, unless you explicitly opt in to training contribution.

If you opt in to training contribution, those images may be used for future model training. Already-trained model weights cannot be untrained image by image.

Deleting your account removes your identification-system data. If an image was already used for training before revocation, that history in model weights is a separate limitation and I want to be clear about that.

Will the AI replace my judgment? No. It is a guide, not an authority.

Can the AI identify hets? No. Hets are genetic carriers that do not express visually.

What if the AI is wrong? It will be sometimes. Low confidence means take the result lightly and verify it yourself.

Who owns the images I upload? You do. ReptiDex only has the limited rights needed to process them for identification, and training use requires a separate opt-in.

Is this for surveillance or ID matching? No. It classifies visible morph traits from photos. It is not an identity system.

Is AI always terrible for the environment? No. The impact depends heavily on the type of model, how often it runs, and how efficiently it is hosted. This tool is intentionally narrow, not an always-on general AI system.

Why trust a breeder-run tool over a generic AI product? Because this tool was built for breeder use, with breeder-labeled data, by someone who uses it on his own animals.

I am Dusty. I am a working crested gecko and leachianus breeder, the founder of Geckistry, and I have spent more than nine years building software professionally. I also hold AWS and Azure certifications. The order matters to me: breeder first, engineer second. ReptiDex exists because I wanted a tool I would actually trust in my own collection.

If the model gets something wrong, if you want to contribute training data, or if you have questions about how this works, email dmumphrey@builtbydusty.com.