Two photos of one forearm scored badly. The lighting hadn't changed. The background had, and it filled most of the frame. Cropping to the subject halved the distance, and that still wasn't enough: testers hit the same thing in ordinary use, where light and background move more than skin does. The score is gone from the app now. A number that mostly reports the room is worse than no number, because people believe it.
Bespoke native software · iOS · macOS · visionOS
Apps that keep your data on your device.
An independent software studio run by Alexander Ramirez, building privacy-first applications, offline machine learning pipelines, and spatial computing prototypes. The work runs on-device by default, no accounts, no servers, nothing uploaded, because the things worth tracking closely are usually the things you'd rather not hand to somebody else.
Available Relocating to Orlando, FL · open to full-time engineering roles and select contract engagements · view opportunities
Four findings from building the apps below. None are benchmarks or marketing figures. Each is a number that came back wrong, or came back surprising, and changed what got built next. Two of them deleted a feature.
A classifier reporting 86% confidence was correct 66.2% of the time, measured across 1,946 held-out observations the model had never seen. Softmax forces ten classes to sum to one, so a photograph of something it was never trained on still comes back as a confident answer. Temperature scaling fitted at T = 1.53 closed an 18.9 point gap down to 1.3. The number was never a probability. It was a ranking wearing a percent sign.
Against 346 held-out photographs of genuinely dangerous spiders, the permissive threshold called 82 of them harmless. Raising the benign floor to 0.86 cut that to one. A second agreement model caught the last one. Getting to zero cost 15 harmless spiders a clean answer, which is the trade worth making: the price of caution is inconvenience, and the price of reassurance is a child reaching toward a recluse.
Asked how sure it was, the on-device language model said 0.67 when it was right and 0.63 when it was wrong. Four hundredths is not a signal you can build on, and it killed the obvious design: route the uncertain cases to a bigger model in the cloud. Whatever decides that local is not good enough, it cannot be the model's own account of how it did.
Kehai Care
Describe how you feel in your own words and get a straight answer about how quickly you should be seen. The part that decides is not the language model. A small on-device model does one narrow job, turning what you said into symptoms drawn from a fixed list of fifty-three, and then shows you what it heard so you can correct it before anything acts on it. The actual decision is ordinary code following the Emergency Severity Index: the most serious thing you reported wins, and readings from the phone can raise the urgency but never lower it. Splitting it that way means the model cannot invent a reason to send you home.
Cairn Skin
A photo journal for watching one spot on your skin over time. Your last photo appears faintly over the camera so you can line up the same shot, week after week, and the whole record exports as a dated PDF you can put in front of a clinician. It does not score the change and it does not tell you what the change means. Both of those turned out to be claims a phone camera cannot support, and neither is the app's to make.
FloraFang
Point your camera at the spider in the garage or the plant in the yard and find out whether it can hurt you. Every app in this category will name a species at 40% confidence. That's fine for a wildflower and reckless for a spider in a child's room, so this one answers a narrower question it can actually support: not what species is this, but is this one of the groups that matters. When it can't tell, it says so and tells you what to do instead, which is a result rather than an error. If something has already been eaten, a separate screen skips identification entirely and routes to poison control, then collects what they will ask for while you're on hold.
Pip: Gift Lists
Every wish list for everyone you shop for, in one place, with real prices. Type a wish the way it was said, “the lego race car with lights”, and Pip finds the actual product at a trusted store, picks the best match for what’s left in the budget, and links straight to the store. It reads a kid’s handwritten Christmas list and the screenshot of a text where someone says what they want. The understanding all happens on the phone; the only thing that leaves it is the product search, with no names or lists attached. Pip himself is a small gift box that changes with the seasons and turns into an elf for the holidays.
Peelback Mechanic
A check engine code tells you a circuit is unhappy. It does not tell you which part of the engine that is or what is buried on top of it. This one takes the code and peels the engine apart on screen, layer by layer, down to the components that code actually implicates. Scoped deliberately narrow to start: one engine, the Toyota 1GR‑FE V6, with a hand-checked map of seventy-five components against forty-three fault codes. A shallow database covering every vehicle would be easier to build and worth nothing in a driveway.
ClearChart
Patient information placed around a clinician instead of on a screen they have to turn away to read. Voice dictation sorts what's said into the right part of the note. Built to explore whether spatial computing does anything real for the documentation load that drives clinician burnout.
B.E.T.H.
A voice assistant running on Apple's on-device model, with no network involved. Alongside it, a test harness that runs a fixed set of sentences through the model repeatedly and reports how reliably speech becomes a structured action. Repeatedly is the operative word: running it across trials, and then across two OS builds, is what turned a clean-looking number into the more useful finding above. Short for Basic Everyday Task Helper.
For Engineering Teams
iOS, macOS, and visionOS developer · Orlando and remote
I am a computer science doctoral student looking to join an experienced team, in Orlando or remotely. I care about clean architecture, responsive frame rates, and native software that respects the people using it. Everything on this page runs on-device by default, and the findings above are the kind of work I want to keep doing on a team that reviews it.
- Swift
- SwiftUI
- SwiftData
- CoreML
- Apple Foundation Models
- RealityKit
- Metal
- Swift 6 Concurrency
- Xcode Instruments
- Git
For Independent Ventures
Bespoke prototypes & on-device engineering
Need a visionOS spike, an offline CoreML pipeline, or a specialized native tool built from scratch? The studio takes on select architectural, clinical, and utility prototypes, and ships them the whole way to the App Store when that is what the work calls for.
- Rapid Prototyping
- visionOS Spikes
- On-Device CoreML
- Offline-First Architecture
- App Store Delivery
Ramirez Labs is Alexander Ramirez. It started with ClearChart, and the question of how to put clinical information in front of someone without burying them in it. Everything else on this page came out of that one, in order.
Partway through it I needed a way to watch a skin change over several weeks, and there was no good way to do it, just a camera roll and a bad memory. The photo comparison work already sitting inside ClearChart turned out to be most of the answer. That piece came out, became Cairn Skin, and is now live on the App Store. What it's teaching me about comparing two photographs honestly is meant to find its way back into ClearChart eventually.
FloraFang followed from that: the same camera, the same problem of an app claiming more certainty than it has, pointed at whatever's in the grass instead of at you. Kehai Care is the same argument again, made about language rather than photographs, and Peelback Mechanic is the one that isn't about health at all, which is mostly a relief. Underneath all of them is B.E.T.H., older than the rest by about twenty years, the talking computer I wanted as a kid and couldn't have until the model started running on the machine itself.
Systems that handle personal information tend to collect more of it than they need, so these apps collect none, no account, no cloud, no analytics. It's a constraint worth designing around rather than apologizing for.
I'm also a doctoral student in computer science, looking at how well small on-device language models turn ordinary speech into reliable actions. The measurements above come out of that work.
Open to iOS, macOS, and visionOS work, especially in healthcare, simulation, and training. Relocating to Orlando. Source for most of this, including this site, is public.
- ResumeDownload Resume (PDF)
- LinkedInAlexander Ramirez
- Emailsupport@ramirezlabs.app
- Codegithub.com/alexanderramirezdev
- BasedAlbuquerque, New Mexico · relocating to Orlando