Teaching an AI twin to practice like its doctor.

What a twin is

A doctor’s digital twin is their avatar and voice clone, working as their assistant: booking, coordination and post-consultation care. Under it sits an intelligence layer that learns how that individual doctor practices: the questions they ask first, how they explain things, what they always check. Their style and decision logic, not medical facts.

The long-term idea is bigger. If regulation ever allows AI to consult, a twin could consult on the doctor’s behalf with the real doctor in the loop. That is future tense, and it is exactly why the foundation has to be built carefully now.

Starting small, on purpose

Twins are live for the 25 doctors in UAT. Each starts from a baseline knowledge base and behavior profile, plus a short setup of at most seven multiple-choice questions that set the twin’s first rules.

How a twin goes from generic to personal

1. 01 · Baseline Knowledge base and behavior profile The same starting point for every twin.
2. 02 · Setup At most seven questions Multiple choice, to set the first rules.
3. 03 · Pilot 4 to 5 trusted patients People who already have a good relationship with that doctor.
4. 04 · Distill SME knowledge, per specialty From sessions with 25+ doctors.
5. 05 · Align The doctor trains the twin Through the methods under test below.

every approved correction becomes a rule the twin keeps

The small patient pool is a deliberate autonomy limit. It is a controlled first pilot: if the twin gets something wrong, it gets it wrong with someone who will tell us.

Expert knowledge distillation

SME sessions with 25+ doctors, across specialties such as cardiology, ENT and pediatrics, feed expert knowledge into the twins. The sessions cover onboarding and product feedback, the workforce agents, and building a knowledge base for each specialty.

The plan is to have five SME doctors per specialty each re-test that specialty’s knowledge base, so it is refined by several doctors and does not inherit one doctor’s bias. Feedback ships in weekly sprints.

The experiment: four ways to train a twin

The product question I care most about: can any doctor train their own twin, without the Pranik team in the room? Doctors are busy, and four training approaches is three too many. So I am testing four gamified, in-app methods for interactive alignment, to pick the best one or two, or a mix, for the production app.

Four training methods under test

Method 01

Consult AI patients

• ·Video consults with AI patient personas, at three difficulty levels; the twin drafts style rules the doctor approves
• +Structured and repeatable
• −Role-play can feel artificial

Testing

Method 02

Quiz your twin

• ·The doctor asks anything and corrects it in plain words; each correction becomes a rule
• +Fast, natural, doctor in control
• −Coverage depends on what they think to ask

Testing

Method 03

Learn from real consults

• ·With consent, patterns across consultations become questions the doctor answers
• +Near-zero effort for the doctor
• −Needs consent, volume and careful privacy design

Testing

Method 04

Role reversal

• ·The doctor plays the patient, then marks up the replay where the twin went wrong
• +Shows blind spots vividly
• −The doctor has to play a convincing patient

Testing

There is a nice symmetry with our evals loop. There, a doctor’s edit of “stomach ulcer” to “peptic ulcer” is not an error; it is a style preference. Here, that same signal becomes a rule the twin keeps.

How I’m judging it so far

The work splits cleanly. Data science builds the models that extract rules and find patterns across consultations. I own the training product around them: the methods, the game, and the decision about which ones survive.