How the AI works
Why a model at all
A lookup table can match “Atorvastatin” to a record. It cannot read “atorvastatan” off a crumpled label in bad light, match it to the medication it means, and then explain in plain language what that medication is for and which side effects a person should know about. That is language and vision work.
The harder half is explaining well. Writing one or two sentences that help an older person understand what a medication is for, without condescension and without a dose, in their own register, is a writing problem. That is the part only a language model does reliably across the long tail of real medications.
So the model is the engine, and betting the product on one lab would be a mistake. Everything above this layer asks for a task, not a model. This file decides which model answers. Changing a route is one line; adding a lab is one case in the dispatch function.
The pipeline
When you snap a label, a vision-capable model reads the primary medication name from the photo. It returns only the name, never a strength, never a dose, never directions. If it cannot find a medication name, it says so. The photo is not stored.
The balanced model reads the medication name and produces a plain-language card: what the medication is generally for, the common side effects worth knowing, and a few practical notes. It is given an explicit system instruction that it must never state a dose or an amount, and the output is scrubbed in code before it reaches you.
As part of the explanation step, the model suggests which reminder slots are typical for that kind of medication from a fixed set: morning, midday, evening, bedtime. It cannot return anything outside that vocabulary. The schedule is reminder times, never amounts.
On Plus and Family, a separate pass reviews your medication list for pairs worth raising with a pharmacist. Each flag names the two medications and gives one plain-language sentence on why it is worth asking about. Only pairs from your actual list are flagged, and any sentence that reads like a dose is removed before you see it.
Routing
This is not a diagram somebody drew. It is rendered from the same routing table the medication engine reads at runtime, so if it is wrong here it is wrong in production.
| Task | Model | Provider | Tier | USD per M tokens |
|---|---|---|---|---|
| Explain a medication in plain language | gpt-4.1-mini | openai | balanced | $1.60 |
| Read the medication name off a label photo | gpt-4.1-mini | openai | balanced | $1.60 |
| Flag interactions to review with a pharmacist | gpt-4.1-mini | openai | balanced | $1.60 |
| Quick routing triage | gpt-4.1-nano | openai | fast | $0.40 |
Above roughly 20,000 characters, a long medication list or an unusually large label image, the request escalates to the frontier model. More context needs more reasoning capacity and the cost difference is worth it.
If the routed model fails or returns nothing, the call walks a chain of candidates from other tiers and other labs before giving up. One provider having a slow afternoon should not cost someone their medication explanation.
Every free-text field the model returns is run through a dosing scrub before you see it. Any sentence that reads like a dose or an instruction to change a medication is removed and counted. The model does not get a vote on this step.
The bright line
What Dosely is
Dosely is information and reminders, not medical advice. It never changes your dose. Confirm anything about your medications with your pharmacist or doctor.
The model is given a system instruction that it must never state a dose, an amount, or an instruction to start, stop or change a medication. That instruction is in the prompt. The dosing scrub is in the code. A false positive drops a sentence, which is safe. A false negative would let a dose through, which is not. The scrub is the medication equivalent of a safety net.
The reminder schedule is constrained to a fixed set of four words: morning, midday, evening, bedtime. The model cannot return a number, a frequency, or an amount in that field. The schedule is reminder times only, never a dose.
Interaction flags are only shown when both medications named are in your actual list. The flag is always framed as a question to bring up with a pharmacist, and the concern text is scrubbed the same way. The count of anything removed is surfaced on the card so you know the guardrail ran.
Emergency guidance
If you have concerning symptoms, a suspected overdose, or a reaction that worries you, get help right away. In the United States call Poison Control at 1-800-222-1222, or call 911, or go to the nearest emergency room.
Candidates
Every model below sits behind the same interface. Adding a lab is one case in one file, which is the point of building it this way.
openai · 1,000,000 token context
openai · 1,000,000 token context
openai · 1,000,000 token context
anthropic · 200,000 token context
google · 1,000,000 token context
meta · 128,000 token context
Today the OpenAI and Anthropic adapters are written and the OpenAI one is live, including the vision endpoint for label reading. The other providers are declared with their real model names and switch on when their key is present. No model is trained on your medication data, by us or by our providers.