Isarar Siddique

Researcher who ships. Founder who publishes.

I build medical AI that actually reaches patients.

Most medical AI dies in the gap between a published result and a working product. Researchers hand off and move on. Founders ship without evidence. I work both sides of that gap on purpose. Six papers open, four companies running, and one screening system live inside the ENT and Neurology departments of a tertiary hospital in India.

See the work →
Isarar Siddique
Dementia screeningAudiometry hardwareDermatology AI Corneal imagingCancer registriesDrug interactions Wearable signalsPCB design Dementia screeningAudiometry hardwareDermatology AI Corneal imagingCancer registriesDrug interactions Wearable signalsPCB design
4Companies running
11Research labs
2Patents filed
6Papers in the pipe

What I run

Four companies solving one problem from four directions.

01

Zero Dementia In trial

Founder and CEO, Biotech Wallah Private Limited

Screening for early cognitive decline on a phone that already sits in someone's pocket. No MRI. No lumbar puncture. No neurologist in the room. Four signals in one sitting, fused into a single risk score in under fifteen minutes.

ONE SITTING, ABOUT 15 MINUTES Voice, read passage Facial expression Fine motor tapping Cognitive battery Featureextraction per signal Late fusion keeps eachsignal separable Risk scorewith reasons Clinicianacts or refers
Late fusion instead of early, so the output can name which signal drove the score. A clinician who cannot see why will not act, and a screening tool nobody acts on is decoration.

The trial spans two departments because the biology ignores org charts. Hearing and speech markers sit with ENT. Cognitive and motor markers sit with Neurology. Both arms are close to done.

Nobody warns you that the model is the easy part. The work was consent infrastructure, recruitment flow, and pulling clean signal out of a room with bad light and a fan running.

  • Multimodal fusion
  • Speech features
  • On-device
  • Genesis 2.0 grant
  • DPIIT DIPP217345
02

Neurento Medtech Building

Co-founder and Technical Advisor, Neurento Medtech Private Limited

An audiometry screening headphone. Right now a hearing test needs a sound treated booth, a calibrated clinical audiometer and a trained audiologist in the room. That stack is why most people in India never get tested until the loss is already bad enough to notice.

We are pushing the calibration into the headphone itself, so the infrastructure cost mostly disappears and screening can happen in a general practice room, a school, or a camp.

HOW IT WORKS TODAY Soundbooth Clinicalaudiometer Trainedaudiologist fixed room, high capital cost, patient must travel WHAT WE ARE BUILDING Calibratedheadphone Phone app guided protocol Audiogram plus referral flag Any room clinic, school, camp booth and dedicated room removed
The hard part is calibration. A clinical audiometer is trusted because its output level is known at every frequency. Moving that guarantee into a headphone is the whole engineering problem, and it is the only reason the booth can go away.

This is also why I learned KiCad, Fusion and LTspice. If you want a better signal then at some point you have to build the thing that acquires it instead of cleaning up after somebody else's sensor. It pairs directly with the ENT arm of the dementia trial, since hearing loss and cognitive decline travel together more often than either field likes to admit.

  • Audiometry
  • Transducer calibration
  • KiCad
  • LTspice
  • Embedded
  • Fusion
03

SuppliAi Live

Founder, Convolity AI Private Limited

Supplement advice is broken because it ignores the only two things that matter. What else you are taking, and who you are. SuppliAi builds one representation of a person from lab panels, prescription scans and wearable streams, then reasons over interaction pathways against it.

INPUTS Lab panels Prescription OCR Wearables Oura, WHOOP, Watch One patientrepresentation Interaction graph5 CYP450 enzymes Docking pipelineDiffDock, DeepDTA Evidence layerPubMed, Lexicomp Rankedguidance
Everything gets cross checked against Natural Medicines, Lexicomp and FDA MedWatch before it surfaces. Concordance sits above 85 percent on the validated set. Anything under that threshold stays hidden instead of shipping with a disclaimer nobody reads.

There is also a longitudinal biological age model scoring six body systems over time. It is the piece I trust least and find most interesting. Composite ageing scores are trivial to build and brutal to validate.

  • Chain of thought reasoning
  • Molecular docking
  • Supabase, RLS, pgcrypto
  • Time series fusion
04

Healoncal Live

Co-founder and Chief AI Officer, Healoncal Private Limited

Skin assessment from a phone camera. The constraint that makes this hard is not accuracy on average. It is accuracy evenly. Most skin models learn on light skin and quietly fall apart on darker tones, which makes them worst exactly where they are needed most.

CONSUMER PHONE, UNCONTROLLED LIGHT Phone image User context Normalisationlight and colour Vision languagemultimodal read Fairness gateacross tone groups Per metricreasoning no bare score
The fairness gate is a gate, not a report at the end. If a metric cannot hold accuracy across tone groups it does not ship. Every number that does ship carries its reasoning with it.

My half is model architecture and getting laboratory grade models onto ordinary handsets without the accuracy quietly leaking away on the way down.

  • Vision language models
  • Explainability
  • Bias evaluation
  • Mobile inference

The question everyone asks

How do you run four companies at twenty two?

The honest answer is that it does not feel like four. It feels like one problem wearing four different shirts.

Zero Dementia taught me how to fuse messy signals from a cheap sensor into a decision a doctor will trust. SuppliAi is that same fusion problem with labs and wearables instead of voice and video. Healoncal is the same explainability problem pointed at skin instead of brains. Neurento builds the sensor. Solve one properly and the other three move forward without opening a second tab.

People keep telling me breadth is a red flag. I think focus is what you recommend to someone who has not found the overlap yet.

I know what my ceiling feels like because I have hit it before. Eleven research labs while finishing a degree, cold emailing every one of them from a college with no research pipeline. That was full. This is not full. I have room, and I would rather spend it than protect it.

Here is exactly how I got into those labs, including the fifty emails that got me nothing first.

Research

Work that stays behind closed doors.

05

Corneal endothelium analysis Private

Ophthalmology, provisional patent filed in India

Cell morphometry straight off specular microscopy. Density, polymegathism and pleomorphism. Those three numbers decide whether a cornea is fit for transplant, and today somebody counts them half by hand.

MICROGRAPH TO SEGMENTED MOSAIC enlarged cell Segmentationper cell borders Cell densitycells per mm2 Polymegathismvariation in area Pleomorphismpercent hexagonal Report
The bottleneck was perception, not classification. Get the cell borders right and all three numbers fall out of arithmetic. Most of the effort went into an annotation plan rather than a bigger model.
06

NextGen OncoData Registry Private

Universiti Malaya, AIM-C01-2025, with Prof. Dr. Sarinder Kaur Dhillon

A national cancer registry rebuilt on WHO ICD-11 instead of ICD-10, covering more than 9,300 records across ten plus cancer types. Registries are unglamorous and they are the ground everything else stands on. No registry, no cohort, no model.

SOURCES Hospital records Pathology Imaging reports ICD-11 codingand anonymising Registry 9,300+ records10+ cancer types Surveillancecross institution ML trainingcohort export
ICD-11 has no clean crosswalk from ICD-10 for several oncology categories. Deciding what happens to the ambiguous ones is a clinical judgement, not an engineering one, which is why a human sits in the coding step on purpose.

Also moving

  • Cardiac signal analysisProvisional patent filed in India. Basis of a proposed collaboration with Khalifa University.
  • Wearable recovery dynamicsCapacity limited physiological resilience, with Prof. Mohamed Elgendi. Paper in revision.
  • Retinal health modelsOphthalmic imaging. Private.
  • Spaceflight brain transcriptomicsFull re-analysis of NASA OSDR cerebellum and hippocampus RNA-seq.
  • Clinical translation LLMsHindi, Indonesian and Malay, with Prof. Irene Li at the University of Tokyo.
  • HealMed datasetMedical data manager and evaluator. Indian, Malaysian and Igbo clinical notes, annotated with practising clinicians.

Code

Twenty something repositories. Most of them are private for a reason.

Screening and neuro

clear-mind-ai

Zero Dementia web client. Assessment flow, session handling and the clinician facing result view.

TypeScriptPrivate

zerodem

Mobile capture app for the screening battery. Records the voice, video and tapping tasks on device.

DartPrivate

alzheimer_ai

Model work for the neurodegeneration side. Training and evaluation for the cognitive risk models.

PythonPrivate

Neuro-AI

Earlier public front end for the neuro assessment interface.

TypeScriptPublic

neuro-ai-backend

Inference and scoring service behind the neuro front end.

PythonPublic

Ophthalmology

corneal-endothelium-analysis

Automated endothelial cell analysis from specular microscopy. Density, polymegathism, pleomorphism, plus an accuracy audit and annotation plan. Patent filed.

PythonPrivate

RetinalHealthAI

Retinal imaging models for disease detection from fundus images.

PythonPrivate

Cancer informatics

cancer-registry-platform

The Universiti Malaya Medical Centre cancer registry platform. ICD-11 coding, anonymisation and multi institution governance.

TypeScriptPrivate

registery-ai

AI layer over the registry. Coding assistance and record normalisation.

TypeScriptPrivate

REGISTRY-ML

Model training and cohort export against registry data.

Private

cancer-back

Registry API and data services.

PythonPublic

cancer-front

Registry web interface for data entry and review.

TypeScriptPublic

onco-data-hub

Oncology data aggregation layer.

TypeScriptPublic

Computational biology

OSD-561-562-brain-transcriptomics

Full re-analysis of NASA OSDR spaceflight brain RNA-seq. Cerebellum and hippocampus, PCA, all contrast DEG landscape, extracellular vesicle and neuroinflammation panel.

PythonPrivate

AfDesign-PeptideBinder

8-mer peptide binder against 5-hydroxyisourate hydrolase (PDB 2H6U_A) using AlphaFold based AfDesign, with pLDDT analysis and mutation sampling.

JupyterPublic

SysBio-DiseaseNetwork

Comparative protein interaction networks across Alzheimer's, type 2 diabetes and NAFLD. Hub bottleneck genes APOE, INS and LEP by centrality and clustering.

JupyterPublic

nanoform-ai

Nanoformulation modelling work, following on from the fermentation paper.

TypeScriptPublic

Platforms and hardware

care-flow

Clinical workflow and coordination tooling.

TypeScriptPrivate

cdce

Clinical data capture and extraction service.

PythonPrivate

aura

Product surface for the personalised health work.

TypeScriptPrivate

med-platform-pro

Medical platform scaffold used across several of the clinical products.

TypeScriptPublic

iit-bhubneshwar-redmud

ESP32 sensing rig tracking NPK, pH and moisture in highly alkaline red mud soil, built at IIT Bhubaneswar. My first hardware project.

PythonPublic

interview

Meeting copilot backend. Transcription, summarisation and session state.

TypeScriptPublic

Private repositories are private because of clinical data agreements, patents in progress or unpublished results. Ask and I will walk you through any of them directly.

Bring me a hard one.

Health AI, clinical informatics, medical devices. Or you are a clinician who thinks one of these models is wrong, in which case I especially want to hear it, because that conversation is worth more than another paper.

isararsiddique@gmail.com