Elicit, then reason.
Ask useful follow-up questions. Connect what the patient shares to the differential, examination and management plan.
Medical education / Research in progress
Practice for
the moments
that matter.
A practice copilot for medical education.
Developing clinical reasoning, patient communication, and empathy through virtual patient encounters.
अभ्यास abhyaas / practice, through repetition
01 / listen beyond the words
Patient / a moment of uncertainty
“मैं…
ठीक हूँ।”
Trainee
“अच्छा… थोड़ा और
बता पाएँगे?”
Illustrative exchange · no audio recording
Clinical reasoningIndic-language encountersEvidence-linked feedbackProgressive practice
01 / The idea
High-stakes skills.
Low-stakes practice.
A good consultation asks more of a doctor than choosing the right answer.
It means following a small clue, making sense of an uncertain history, weighing the next step, and hearing the concern behind a patient’s words.
Med Abhyaas explores a place to practise those skills, make mistakes, receive specific feedback, and try again before taking on comparable responsibilities in patient care.
Listen, then clarify.
Explore tone, hesitation and implicit concerns sensitively. A cue is an invitation to ask, not proof of an emotion.
Reflect, then return.
Review the exact moment. Rehearse with support, then revisit the skill in a new case without hints.
02 / Everyday language, clinical meaning
Patients don’t speak
in textbook chapters.
Explore three constructed teaching examples. These are interface previews, not live AI consultations or validated clinical cases.
Training mode / illustrative exchangeScripted preview
Patient
“खाँसी तो है… और कपड़े भी कुछ ढीले लग रहे हैं।”
“There’s the cough… and my clothes seem a little looser.”
A possible follow-up
“अच्छा… वेट, मेरा मतलब वजन, कम हुआ है? कब से ऐसा लग रहा है?”
“Okay… weight, I mean your weight, has it gone down? Since when have you noticed this?”
What the learner can practise
Notice the indirect clue, characterize it, and use the clarified history when considering the next step. The clue alone does not establish a diagnosis.
Regional wording requires native-speaker and clinician review. Language breadth and expressive audio are development targets; this preview does not demonstrate either.
Listen / a 30-second synthetic walkthroughOne encounter.
One encounter.
Both sides in Hindi.
Both sides speak Hindi, with draft Haryana-region wording for the patient and terse, matter-of-fact Hindi questions from the trainee. English is a reading aid, not the spoken default.
Read the clip transcript and voice limitations
SP/VP: डॉक्टर साब... खाँसी सै... कई दिन हो लिए। थोड़ा चलूँ तो... साँस फूल जावे। घबराहट हो री सै।
MT/SD: खाँसी कब से है? बैठे हुए भी साँस फूलती है?
This synthetic draft uses two distinct Sarvam Bulbul v3 voices: Ritu (SP/VP) and Shubh (MT/SD). The input uses regional wording with Hindi selected. The background assessment and five-domain progress ratings are prewritten, fictional illustrations, not live inference or real learner data. Sarvam’s documented API has no dedicated Haryanvi selector; authentic accent, clinically realistic respiratory sounds and distress expression remain unvalidated. Literacy is a separately authored attribute and cannot be inferred from this voice.
03 / An inspectable architecture
From a guideline
to a learning moment.
Two connected flows: author and approve the knowledge; retrieve only what applies. Then keep the patient’s clinical truth fixed throughout the encounter.
The internal problem representation is an evolving clinical summary. It may use English, while the patient and trainee speak in the selected Indic language with code-mixing. Keep the original speech linked to every interpreted fact.
Read Robert Wachter on problem representation in “The First Problem with the Argument: The Reliable Fact-Set and GIGO.” A clinical summary belongs to the learner/examiner pathway; the patient Talker receives only permitted patient knowledge.
Processing Human / policy gate Versioned stateProposed design · select any stage
A
Author & approve
Before the encounter Only published, versioned knowledge enters the retrieval flow.
B
Retrieve & apply
Scoped to a case and purpose The case supplies patient-visible facts. The examiner receives the scoped rubric.
Fast conversation path
Trainee ⇄ Patient Talker
Coordinator → disclosure policy → permitted reply → delivered audio
Asynchronous, restricted roles
Planner + Perception
English internal state · Indic-language interaction · attributed cues · separate examiner problem representation
Separate evaluation path
Evidence → review → practice
Fixed exam / order results · silent examiner · faculty correction · learner feedback
Source changes or revoked claims trigger impact review and a new release. Active cases are never silently rewritten.
Stage 01 / Author & approve
Authorized sources
Start with documents the institution is permitted to use. Record publisher, edition, effective date, access rights and a checksum. Source ownership and processing permission stay attached to the knowledge.
Boundary to preserve
A publicly readable guideline is not automatically licensed for redistribution or model processing.
Read the complete flow as text
Authorized documents are normalized, then clinical recommendations are extracted with source spans and their conditions. Clinicians review the interpretation. A publication gateway checks approval, permitted use, freshness and conflicts before creating an immutable release.
A scoped request retrieves candidates only from an approved release. The applicability gateway checks the case population, setting and clinical conditions. It returns applicable evidence, a coverage gap or an explicit unknown. Faculty review the case and rubric before pinning a case version for the encounter.
The patient Talker receives only permitted patient facts. The Planner and Perception Observer work asynchronously with restricted state. Fixed examination and investigation services return authored findings. A separate examiner links delivered evidence to reviewable feedback. Corrections update learner recommendations; source changes require impact review and a new release.
04 / Practice that progresses
Help now.
Independence later.
The aim is learning that survives a delay and transfers to a new patient conversation. Finishing a case with hints is useful practice, but it is not the same as independent performance.
- 01
See an example
Understand a useful question or clinical decision.
- 02
Try with support
Use graduated hints and focused rehearsal.
- 03
Practise independently
Apply the skill in a different encounter.
- 04
Return after a delay
Revisit the skill and examine what transfers.
05 / Evidence & development
A research project.
Built to be examined.
Simulation fidelity, clinical correctness, assessment reliability and learner benefit are separate questions. Each needs its own evidence.
Controlled case mechanics
A local demonstrator uses two synthetic cases, fixed clinical facts, selective disclosure, examination and order lookups, and an event timeline.
Reviewed knowledge + expressive speech
Clinical release gates, scoped retrieval, streaming agents, and validated regional speech are the next development targets.
Independent performance
Live testing is reported as ongoing. No participant outcomes, validated competency scores or clinical-transfer results are presented here.
Ideas informing the project
01 / Oncology education · preprintOnco-ShikshakIllness scripts, scaffolding, spaced practice and feedback.↗02 / Multimodal consultation · preprintTowards Expert-level Medical AI for Real-time Video ConsultationsAsynchronous roles and the distinction between perception and action.↗03 / Knowledge infrastructureKnowledge Layer for AgentsA foundation for source-linked extraction and human review, with additional clinical gates proposed here.↗These are related projects, not endorsements or demonstrated integrations. The page is a sample research presentation; the interactive encounter is prewritten and does not collect microphone input or patient information.