How Sourced AI works
It answers questions about type 2 diabetes medications from sources picked by hand, and shows you two things about every point it makes: how strong the evidence is, and whether it was studied in people like you.
The research is agentic. Instead of one fixed search, an agent plans each step: it searches again from new angles, checks who each study enrolled, pulls newer research, and stops to ask you when a missing detail would change the answer. Fixed rules, not the AI, grade everything it finds.
What happens when you ask
Check for anything urgent
Before anything else, fixed rules scan your message for emergency signs. If one matches, you're pointed to help right away and nothing else runs.
Research your question
A research agent searches the library, tries new angles when results are thin, checks who was enrolled in key studies, and looks for newer research and Health Canada notices. If the answer depends on a detail it doesn't have, it stops and asks you.
Grade the evidence
Each source gets a grade from the type of study behind it, by fixed rules. The AI never picks a grade.
Check whether it fits you
Your profile is compared with who was actually in each study. Every mismatch is named, like “Study ages 40–65. Yours: 65–74.”
Answer, or say what's missing
If the evidence is strong enough, you get an answer in plain language with its grades and sources. If not, you're told exactly what would be needed, and nothing is guessed.
How strong is the evidence?
Higher on the ladder means more trustworthy. Each bar is one step.
Diabetes Canada's recommendations, written by expert panels from all the evidence.
Pools the results of many studies on the same question.
Health Canada's approved label: dosing, warnings, side effects.
Enough people, randomly assigned, to trust the result.
Randomized, but too small to be sure.
Follows people over time. Shows links, not cause.
Useful for ideas, weak for decisions.
A source drops a step when it has a known weakness, like a high risk of bias. A trial counts as large when it was designed with enough people to detect the outcome in question, or, where that isn't recorded, enrolled at least 300 people.
Does it fit you?
- Studied in people like you
- Partly matches people like you
- Studied in people quite different from you
- Couldn't fully check against your profile
- General guidance, not a single study
“Couldn't fully check against your profile” means a detail is missing from your profile or the study didn't report it. Your profile is where you fill these in.
Newer research and safety notices
The agent can also search PubMed for research from the last five years and check Health Canada's recalls and safety alerts. Those results appear in their own panel marked not yet reviewed. They never change the answer or its grades, because nobody has checked them or who was in those studies. Studies that come up often are reviewed by hand and may be added to the library.
What's in the library
Every source the main answers can draw on, strongest first.
Loading the library
What it won't do
- Diagnose anything, or tell you to start, stop or change a medication.
- Answer about medications outside the library. Today that's metformin and GLP-1 medications.
- Replace your doctor or pharmacist. It's built to help you bring better questions to them.
If you think you're having an emergency, call 911. For health advice in Alberta, call Health Link at 811.
Who made this
Hi, I'm Akhil.
I'm a computing science student at the University of Alberta, specializing in AI. I built Sourced AI after watching someone in my family manage type 2 diabetes.
Every new prescription came with questions a short appointment couldn't cover. Searching online gave confident answers that never said how strong the evidence was, or who it was tested on.
So this does the reading. It searches guidelines, drug labels and trials, grades the evidence, checks whether it was studied in people like you, and tells you when it doesn't know.
Akhil Vishnubhotla · Edmonton, Alberta