Clinical AI Diagnostics for POC & Last-Mile
Normalize Every Imaging Device. Power Clinical AI Anywhere.
HbPRISM measures haemoglobin and bilirubin non-invasively from a smartphone camera — with clinical-grade consistency made possible by Norma Engine, the normalisation layer that turns any consumer device into a calibrated clinical instrument. CDSCO-targeted. Community health worker deployable.
The Problem
Anaemia, jaundice and neonatal hyperbilirubinaemia remain major public-health challenges. Yet objective measurement is often constrained by blood draws, laboratory infrastructure, trained personnel and -when using imaging AI- variability across devices.
Anaemia affects hundreds of millions worldwide, with particularly high burden among children and women. Reliable haemoglobin measurement remains difficult to access at the last mile.
Objective bilirubin measurement requires laboratory testing, creating barriers to timely assessment where diagnostic infrastructure is limited.
Newborns require timely identification of clinically significant hyperbilirubinaemia, yet access to objective bilirubin assessment remains uneven; particularly outside well-equipped facilities.
The populations that need screening most are often furthest from laboratory infrastructure. What is missing is an objective diagnostic layer that can operate at the point of care.
Smartphones and edge imaging devices vary in sensors, ISPs, firmware, computational photography and illumination- changing the image before clinical AI ever sees it.
When device variation is mistaken for biological variation, AI performance degrades. Clinical AI needs an infrastructure layer that standardizes its inputs.
Founding Story
In 2025 our founder watched a clinical AI diagnostic get pulled at pilot — not because the science failed, but because every smartphone camera sees colour differently.
That is why Gentrac Labs exists: so image-based clinical AI can survive contact with the real world. We built NormaEngine, a device-agnostic image normalisation layer, patent pending, that makes smartphone screening reliable on any phone. On top of it: one-minute, non-invasive anaemia and jaundice screening - no blood draw, no lab, no biowaste - built for frontline health workers across India and other low- and middle-income countries. It is designed for the people least likely to reach a laboratory: newborns, children, pregnant women and the elderly. Febrile illness triage is next.
Proof of concept is complete. The regulatory path is scoped. Institutional partnerships are active.
"If you fund, deploy or study last-mile health, lets talk."
The Insight
Every consumer smartphone applies its own proprietary image signal processing, tuned for photographic aesthetics rather than clinical measurement. The same conjunctiva renders differently on an iPhone and a budget Redmi — and the model answers anyway, confidently and differently. Clinical AI trained on curated device sets fails on the phones patients actually own.
The biometric industry met this problem years ago and solved it. Iris and facial recognition normalise to a canonical representation before feature extraction — deterministic, hardware-invariant signal processing at the input layer, not retraining per device. The approach is codified in ISO/IEC 19794-6 and runs in every compliant system worldwide.
No equivalent existed for clinical imaging, where the target is living tissue rather than a template and the reference has to travel with the patient. NormaEngine is that layer: deterministic correction for camera, illumination, spatial and skin-tone variation, applied before any model runs. Patent pending.
From a single photograph of the eye, HbPRISM estimates haemoglobin from the palpebral conjunctiva and total bilirubin from scleral icterus. Read together they do what neither does alone — separating nutritional anaemia from haemolytic causes such as malaria and G6PD deficiency, and from hepatic causes such as hepatitis. A frontline worker gets a direction to act in, not a number to interpret.
For newborns the measurement moves from eye to skin, where jaundice carries the highest mortality risk and reaches the fewest babies. It is the harder optical problem: skin carries melanin and perfusion as active confounders, where the sclera is avascular and comparatively melanin-poor. Solving it there is what makes screening possible at the bed-side.
Both models are benchmarked on published clinical datasets — haemoglobin at AUC 0.95, newborn jaundice at 93.5% specificity against 92.6% sensitivity. Multi-site clinical studies with Indian institutions are planned.
NormaEngine makes the measurement trustworthy. HbPRISM makes it matter.
The Platform
Three-layers: an open normalization SDK, a proprietary intelligence layer keeping it regulatory grade, and the clinical models that run on top.
Deterministic, device-agnostic image signal normalisation. NormaCard- a disposable calibration reference that travels with the patient — makes the correction per-image rather than per-device, so unseen hardware needs no retraining. Our goal is to democratize smartphone based Clinical AI solutions.
Continuously maintained device profile library, real-time firmware drift monitoring, and a per-image immutable audit trail built for regulatory submission. Making it easy for every SaMD company's regulatory dossier, referenced to a specific Norma library version.
Three CDSCO-targeted models on one SDK. HbNeT estimates haemoglobin from the palpebral conjunctiva. BiliNeT estimates total bilirubin from scleral icterus. BiliGrad recovers bilirubin from neonatal skin, a physics-informed model, where melanin and perfusion are active confounders. One capture, no blood draw, no biowaste.
Traction
Diagnostic Pipeline
Eye-based and skin-based imaging, running on one SDK. Community deployable screening at a fraction of the current diagnostics cost. No blood draw. No lab. No biowaste.
Non-invasive haemoglobin estimation from a single eye photograph. Benchmarked at AUC 0.95, 94% sensitivity, 89% specificity on a published clinical dataset, measured on a held-out split.
Non-invasive bilirubin estimation in newborns from body photography. At matched 92.6% sensitivity, the physics-informed head raises specificity from 87.0% to 93.5% on a clinical dataset.
Total bilirubin from scleral icterus, captured in the same photograph as haemoglobin. Architecture defined; enters first human testing within the clinical programme
Both biomarkers from one capture, separating nutritional anaemia from haemolytic and hepatic causes- triage a frontline worker has not previously been able to perform.
Additional non-invasive biomarkers extend on the same normalisation layer without new hardware, each inheriting the device library, the drift monitoring and the audit trail.
The Team
All four have built, operated, or commercialised healthcare products from the ground up — and one of us has lived the exact failure we are solving.
Scientist-CEO with 25 years translating research into commercial healthcare. Built Tech Care for All — 350,000 HCPs, 280 institutional partners, Rx & MDx -India & Africa ;Telemedicine & EMR. Founder…
Associate Professor in Artificial Intelligence & Data Science at Jio Institute. Specialist in domain generalization, distribution shift detection, self-supervised learning, uncertainty-aware deep learning. Leveraging AI to expedite clinical trials in…
AI researcher with history of delivering enterprise-scale AI products. B.Tech in IT Specializing in: Artificial Intelligence, Computer Vision and Deep Learning Built and deployed enterprise-scale AI solutions across government platforms,…
Insights
On 13th August 2026, Gentrac Labs signed a milestone institutional partnership with PI-RAHI: Northern Region S&T Cluster, of…
This is one of those moments that quietly changes things. Prof. Sudipta Roy has joined Gentrac Labs as…
Beyond the “Black Box”—Building Trust in High-Volume CareLive from Delhi at the Health AI Summit! Consensus among the…
