Early stage · Simulation language model · Open for investment
A language model that simulates the human body.
MedDeck is an AI collaboration with medical colleges. We are building a dedicated simulation language model, on its own architecture, that classifies medical knowledge into a 24-class taxonomy. The aim is to simulate the whole body, so students can meet rare diseases in class, not for the first time in practice.
- 24 knowledge classes
- 24,000+ medicines in the dataset
What we are building
From a simulated body to rare diseases in class
The gap in medical learning
Students learn on clinical rounds, but the cases they meet depend on who happens to be admitted. Rare diseases may never reach them before they graduate.
The gap in medical AI
General language models generate text from an open vocabulary, so they can state relationships that have no verified source. Simulating a body needs a model tied to structured knowledge.
The idea behind MedDeck is a simulation language model that can represent a whole human body, with its systems, conditions, medicines and reactions.
That led to our go-through plan: put simulated cases, including rare diseases, in front of medical students during their normal classes and campus routine. Practising on a case is more practical than reading about it.
- A dedicated modelA simulation language model on its own architecture, trained on text classified by our 24-class taxonomy. In development.
- A knowledge graphThe classes link into a graph, so every concept sits inside a context of related concepts instead of standing alone.
- A medicine datasetMore than 24,000 medicines catalogued with generic and brand names, manufacturers, indications and precautions.
- Rare diseases in classPlanned: simulated cases of rare conditions during regular classes, so students practise decisions on them before they meet one on a ward.
- Go-through planPlanned: a national simulation competition for medical colleges, and simulation stands placed in college halls and corridors, so practice fits into the academic day.
- Medical college collaborationWe are building with medical colleges, starting with knowledge review and teaching use.
For pharmaceutical partners
Where students meet your medicines
We plan to work with pharmaceutical companies. Their medicines enter MedDeck's library, and medical students meet them inside simulated cases that show how each one behaves across diseases, patients and reactions.
Concept, in design
- Every year's cohortMedical students use the library through their training. Usage is planned to be reported by student year, so partners can see which cohorts met their medicine.
- Clinical contextA medicine appears as a choice the student has to justify. They see the outcome and any adverse reaction, so exposure comes from how the medicine is used.
- Partner reportCases that featured the medicine, how often students selected it, and what happened next, including reactions.
- Bacterial pneumonia
- Viral fever ✓
- Appendicitis
- ✓Student recognises the allergic reaction
- ✓Stops the medicine and treats the reaction
- ✓Records the reaction in the patient's notes
Concept preview of a planned feature. Not clinical guidance.
Become a partnerAI and compute
A taxonomy first, then generation
MedDeck is building a dedicated medical model with its own architecture. It starts with a purpose-built 24-class taxonomy that classifies medical knowledge, and links the classes as a graph.
Unlike general entity schemes, the tag set also covers negation, certainty, conditions and grammatical role, so every clinically meaningful word carries a defined role.
- Medical text corpusPlus the 24,000+ medicine libraryImplemented and tested
- 24-class annotationSilver-labeling with Gemma 4 (31B, instruction-tuned), prompt refined over several roundsImplemented and tested
- Inter-class graphTyped links between the classesDesigned, not yet built
- Context worldGraded priority field around a seed conceptDesigned, not yet built
- Restricted generationMulti-hop grounding chains for distant conceptsDesigned, not yet built
The multi-hop grounding requirement is the proposed novel contribution. It is pending implementation and evaluation.
Content classes 13 · what the text is about
SYSWhole body systemsORGOrgans and structuresLOCNon-organ sitesPHYPhysiological processesBIOMolecules, ions, enzymesPATPathological findingsDISDiseases, syndromesSYMReported symptomsINVTests and imagingPHCDrug classesMED_GENGeneric ingredientsTRTNon-drug treatmentsRULEProtocols, guidelinesAttribute classes 6 · reshape a concept
QUALQualitative descriptorsQTYNumerical quantitiesUNITUnits of measurementTIMETimeframes, durationsFREQFrequencySTATEClinical statusLogic classes 4 · invert, scale or gate a concept
NEGNegationCERTDegree of certaintyCONDConditional operatorsROLEFunctional rolesOperational class 1
ACTClinical action verbsWhere GPU compute comes in
- Large-scale taggingRunning the instruction-tuned labeler over a growing medical corpus.
- Embedding trainingA customized medical embedding space built on the 24 classes.
- Fine-tuningTraining a simulation language model on the tagged data.
- InferenceServing responsive patient simulations to students at scale.
Compute for these stages is currently covered through competition-based resources, public platforms and self-managed GPU environments.
Progress
Recognition and roadmap
- Phase 1Computational coreMedical embedding system, stronger concept relationships, better simulation accuracy.
- Phase 2Education accessBring simulated cases, including rare diseases, into medical classes and campus spaces.
- Phase 3National networkExpand across medical colleges and hospital learning programs in Bangladesh.
- Phase 4Shared knowledge graphGrow the graph through contributions from students, researchers and institutions.
Team
Built by three students
- AHAsadul HaqueCo-founder
Data Science, United International University - MRMehedi Hasan RuhiCo-founder
Data Science, United International University - SSSaif Azad SamiCo-founder
Data Science, United International University
MedDeck began as a student project at United International University, Dhaka, and has been developed over the past year.
Open for investment
Invest in MedDeck
MedDeck is an early-stage AI collaboration with medical colleges, open to investors who want to back it from the start.
Investment funds the model, the compute behind it, and the first partnerships with medical colleges and pharmaceutical companies.
Where we are
- 24-class taxonomy designed and in use
- LLM-assisted tagging pipeline built and tested
- 24,000+ medicine dataset catalogued
- Hult Prize 2025 and CSE Project Show 2026
What we are building
- A simulation language model on its own architecture
- Context-world and grounded generation
- Rare-disease cases for students, in class
- Partnerships with medical colleges and pharma
What investment funds
- GPU compute for tagging, embeddings and training
- Medical knowledge development and model improvement
- Pilots with medical colleges
Tell us who you are and we will reply with details.
Your message is ready. Send it to this address from any email app.
meddeck.officials@meddeck.space
- Asadul Haque
asadulhaque@meddeck.space - Mehedi Hasan Ruhi
ruhi@meddeck.space - Saif Azad Sami
saif@meddeck.space
Madani Avenue, Dhaka, Bangladesh