ZenMD Builds Medical AI That Refuses to Guess

August 25, 2026 14 min read

Pranav Reddy - ZenMD Co-Founder and Chief Marketing Officer

Almost every AI company sells you on what its model can answer. Pranav Reddy, co-founder and Chief Marketing Officer of ZenMD, would rather talk about what his refuses to answer. Ask a general-purpose chatbot a medical question and it will nearly always agree with you, confidently, in fluent prose. In a clinic, that agreeable streak is not a feature. It's the failure mode. ZenMD's whole architecture is built around a sentence most AI products are engineered never to say: I cannot verify this.

ZenMD is an India-focused clinical AI platform that Pranav built with his sister, Rithika Reddy, the company's Founder and CEO. It launched in India in May 2026 with four workflows and a verification layer underneath all of them. And it started, as good ideas often do, with someone being wrong for years.

A Marketer, a Fintech Builder, and a Summer in India

Pranav has spent his entire working life in marketing, based in the US. An agency internship became a career, the career became a director's role, and eventually he left to run his own shop without, as he puts it, any guardrails. But the thread running underneath all of it was never advertising. It was human behaviour.

He was the kid taking AP and honours psychology classes, fascinated by mob mentality and why people do what they do. Marketing was simply the commercial expression of that curiosity — behavioural symptoms plus data, on repeat.

Rithika came at it from a different direction: over a decade in fintech, building lending platforms, working with interest rates and the algorithms that sit behind credit decisions. Two siblings, both fluent in extracting signal from messy human data, neither of them a doctor.

The trigger was a family trip. They were in India visiting their grandparents, getting routine checkups done, and they noticed something that anyone who has sat in an Indian OPD queue will recognise instantly.

"We have incredibly smart doctors, but they are constantly on the run."

That observation went home with them and marinated. AI was just starting to get real traction. The question forming was narrow and practical: what could it do to make a doctor's day less impossible?

The Friend Who Was Not Lactose Intolerant

Then a friend came back from India with a story.

The Diagnosis That Started a Company

For years, this friend had been told by doctors in the US that they were lactose intolerant. It was a reasonable conclusion. The symptoms fit, and in an American clinical context, that is the high-probability answer.

On a trip to India, a local doctor identified the actual cause almost immediately: a worm, of the kind you pick up from bacteria-contaminated water.

Same symptoms. Same patient. Completely different diagnosis — because the two doctors had been trained to look for different things.

Pranav's point is not that the American doctors were bad. It's that they were pattern-matching against the population in front of them, which is exactly what good clinicians do. Waterborne parasites simply are not high on the differential in New York the way they are in Hyderabad.

"The doctors there are not looking for the same symptoms as the doctors here. And that's when the light bulb moment happened — what if we connect both sides? You can give the diagnostic power of someone sitting in New York or Miami to someone who is in a village in India."

That's the founding thesis in one line, and it runs in both directions. The village doctor gets access to what a specialist in Miami would consider. The Miami doctor gets a system that knows a tapeworm is a live possibility in a patient who spent a month in eastern Asia.

Why Western-Trained AI Misses Indian Bodies

Ask most founders whether ChatGPT and Gemini are Western-biased and you'll get a confident yes. Pranav gives a more careful answer, and it's better for being careful.

He won't say the models are 100% Western. He says they're skewed, and he's specific about the mechanism. You train AI on documented history. For decades, a disproportionate share of medical literature — inventions, case documentation, trial data — was produced and archived in the West. The large model companies are also based there, so their access to that corpus is structurally easier. It isn't that India has no data. It's that the documented, digitised, machine-readable slice tilts one way.

Which led to the decision that shaped the product: they refused to take an existing Western-trained system and adapt it at the edges.

"We needed something that was ground up to be Indian. It needed the Indian drug doses, Indian food habits, how we put on fat, how we have a high carb diet compared to the rest of the world."

He gets granular about why adaptation fails. A generic system tells a patient to eat more fish. Large numbers of Indians will not eat fish. Others won't eat root vegetables. Others eat only specific meats, or none. Dietary advice that ignores this isn't slightly off — it's unusable, and a doctor who gets unusable advice twice stops opening the tab.

Adapt vs. Build Ground-Up

Adapting a Western model: Faster to ship. Inherits a probability landscape tuned to a different population — different prevalent conditions, different body composition, different drug dosing conventions, different diets. The gaps show up exactly where clinical stakes are highest.

Building ground-up for India: Slower and harder. But the baseline assumptions — carb-heavy diets, vegetarian and religious dietary restrictions, locally prevalent infections, Indian dosing norms — are in the foundation rather than bolted on as an afterthought.

ZenMD's read: You cannot patch a population mismatch with prompt engineering. It has to be structural.

Doctors Test It, but Data Has to Argue With Itself

So how do two non-doctors validate a clinical product? Pranav answers it in two halves — the human layer and the data layer.

On the human side, ZenMD works with specialists across multiple disciplines who continuously QA the output. At the time of the conversation he described covering over 14 specialties with a target of close to 50, with doctors constantly testing whether what the system says is actually relevant in practice. (Company materials published since put the platform's coverage at more than 80 specialties — the validation panel and the coverage claim are different numbers, and it's worth keeping them apart.)

On the data side, the design principle is one that a lot of medical AI marketing quietly dodges.

"There is no right answer, guys. At the end of the day, it's a diagnosis, and we expect the doctors to complete that diagnosis. This is just an assistant. We're not trying to replace or remove any of the doctors."

His analogy is researching something on the internet twenty years ago. The search engine gave you material. You still made the decision. ZenMD is trying to be a much better version of the material, not a replacement for the decision.

Underneath that sits what he calls multi-layer verification. The interesting thing is that the model isn't asked to produce an answer. It's asked a chain of yes-or-no questions, each one checked against something external.

How the Verification Chain Works

1. The clinical question comes in. A doctor asks whether a presentation is, say, a common cold.

2. The system checks cited material. Not a private hoard of proprietary data — Pranav is explicit that they don't have secret data nobody else can reach. The differentiator is what the algorithm does with public, citable sources.

3. Source quality is interrogated. Was this in a reputable journal? Are there actual case studies confirming it?

4. It decomposes into yes/no checks. Does a common cold present with this symptom — yes or no? Each answer opens the next layer of questioning rather than closing the loop.

5. The output carries a confidence level and its citations. The doctor sees not just the conclusion but the reasoning trail and how sure the system is about it.

The Model That Will Never Tell You You're Wrong

This is the sharpest thing Pranav says in the whole interview, and anyone who uses AI daily will feel it land.

"ChatGPT or Gemini in today's day and age is always built to answer you. It's not built to tell you you're wrong. If you've tried it, you know — with any question, it will always say yes. It'll never say, 'Priya, you're wrong.' It's always built to just answer you."

General assistants are optimised to be helpful and agreeable. In most contexts that's harmless, occasionally annoying. In a consultation room, an agreeable machine is an amplifier for whatever the doctor already suspected — which is the precise opposite of what a second opinion is for.

"Don't just answer me — actually validate what I'm asking. Then tell me which sources say it's true. But at the end of the day, leave the answer to me as a doctor. I'll just give you all the supporting documents."

Two things happen when the output is uncertain, and the second one is the product.

Two Modes, and Only Two

Confident: ZenMD returns the answer, the cited source, and an explicit confidence level — here is what I think, here is why, here is how sure I am.

Not confident: It stops. "I'm sorry, I'm not able to verify this, so I would suggest digging deeper." It will ask for more patient information, or a clearer image, rather than filling the gap with plausible-sounding text.

They demonstrate this deliberately. When the team runs live sessions at medical colleges — students use the platform too — they'll upload something genuinely ambiguous just to show what happens.

"It will check itself and say, 'I'm sorry, I cannot proceed unless I'm confident.' And that's the biggest layer, where a generic AI will go in and say, 'Yes, Priya, you are absolutely right, this is exactly what I think it is' — and that's where you go into the rabbit hole. This stops before that door even opens."

That last phrase is the whole safety argument. A hallucination in healthcare is rarely a single wrong sentence. It's the first step of a chain — a wrong anchor that contaminates every subsequent decision. The intervention has to happen before the door opens, not after.

Selling Clinical Software When You're Not a Clinician

Doctors are trained sceptics, and the first thing they check is the source of a claim. Two founders with marketing and fintech backgrounds telling them how to practise medicine is, on its face, a hard sell.

Pranav doesn't dodge it, but he reframes it neatly. Apprehension about a new tool is universal and has nothing to do with who's pitching it — nobody adopts clinical software because a stranger sounded convincing. And on the credentials question, he flips the direction of the argument.

"If I was a doctor, I would have come up with the idea and then got some help to build it. So it's vice versa — we've built something and we are taking doctors' help to refine it. You still need both sides to implement it."

You don't ask why the person building a company isn't also the architect and the engineer. Between them, the siblings bring behavioural algorithms from fintech and post-hoc behavioural analysis from marketing — how people react, what they choose, what they actually do rather than what they say. The medicine comes from the doctors. Neither half works alone.

Trust, Pen and Paper, and Repeat Users

India's clinical culture still runs substantially on paper. Plenty of doctors write prescriptions by hand and rarely reach for a phone mid-consult, let alone a laptop. So what does adoption actually look like?

The metric ZenMD watches hardest is repeat users, and Pranav says it has been climbing steadily. But he's careful about the framing — he describes the company as still at the foundation stage of building trust, not as having won the market.

"It can help them surface possibilities, organise information, reduce repetitive work — but it shouldn't remove the reasoning process that you have a doctor for."

The pitch, on the website and in person, leads with efficiency rather than intelligence: we want to make your day more efficient. And the efficiency case is brutal arithmetic. A doctor in a smaller town might see sixty or seventy patients in a day. In a dense metropolitan area, it can run into the hundreds. Same doctor, same finite hours — the time available per patient collapses.

That's where ZenScribe comes in. If the consultation documents itself, the doctor isn't writing OPD notes afterwards or re-narrating the whole encounter to the attender outside. They're present in the conversation instead of transcribing it.

The Use Case Nobody Predicted

Here's the part founders can only learn by shipping. The team expected documentation to dominate. What actually surged was drug dosage.

Not the obvious kind — every doctor knows a paracetamol dose. The queries coming in were about the modifiers. A patient of a given body weight, with a particular carbohydrate intake — what dose does that actually call for? Someone who is six foot four carries a different mass and absorbs differently from someone who is five foot one. Does an extra ten or fifteen kilos move the number enough to matter?

These are questions a doctor can answer if they stop and think. The problem is that nobody has four spare minutes to stop and think when the queue is a hundred deep.

The Supplement That Binds to Your Morning Coffee

Pranav's favourite example from testing: there is a supplement which, taken alongside caffeine, binds to it and leaves the body — carrying the intended benefit out with it.

The doctor prescribes it and says take it in the morning. Most people drink coffee in the morning. Nothing about the prescription was wrong, and yet the patient gets nothing from it.

"Those are not things you're thinking about when you're seeing a patient for four minutes, five minutes."

This is the honest shape of clinical AI value. Not replacing diagnosis — catching the interaction that a rushed human brain has no bandwidth to surface.

Usage data the company has since published from Hyderabad backs the pattern up. Diagnosis and symptom analysis leads at 21.6% of queries, but drug information and dosing sits second at 15.6% — ahead of treatment planning at 10.9% and lab and imaging interpretation at 9.3%. The surprise held.

Four Workflows, One System

ZenMD ships as four named capabilities, and Pranav is insistent about a distinction that sounds like semantics until you think about it.

The Four Workflows

ZenScribe — real-time voice-to-clinical documentation. The consultation becomes structured notes without the doctor writing them.

ZenNotes — automated OPD notes, the paperwork that otherwise piles up between patients and eats the evening.

ZenConsult — evidence-based clinical decision support. Consultation history, drug dosing, the verified reasoning trail.

ZenScan — medical imaging analysis. Upload a CT, a bone scan, a dental scan, and ask where the anomaly or the plaque is.

"I wouldn't say it's a product, because it's just one big umbrella. These are workflows. You can choose a different workflow to perform a specific task."

The four names exist so clinicians can find things, not because there are four separate systems. Underneath, a single platform is trying to understand where in a workflow the doctor currently is and surface the right information at that moment. Upload a scan and ask about plaque, and ZenScan handles it — but the verification layer, the citations and the confidence thresholds are the same ones running everywhere else. Sold as four products, they'd be four integrations and four logins. Sold as one system with four doors, they're a place the doctor already is. It's a philosophy shared by Nura's AI-powered full-body screening, where the imaging intelligence matters far less than whether it slots into an existing clinical flow.

Separating Who You Are From What You Have

Healthcare data privacy is a serious discipline in the US and a rapidly hardening one in India. And there's a permanence problem specific to AI: feed personal data into a model and you cannot meaningfully pull it back out.

ZenMD's answer starts with a decision about sequencing.

"These are things you have to build in the beginning of the platform. You can't retrofit it later. You can't have half the data be exposed and then say, oh, I'm going to block some of it."

So they built to the strictest regime rather than the local one — HIPAA, DPDP and ABDM compliance from day one, each with its own verification requirements, on the reasoning that a platform built for the loosest applicable jurisdiction can never be upgraded into the tightest.

"This is healthcare, which means it has a lot of scrutiny. So the data deserves a much higher standard than ordinary consumer data."

The architectural trick underneath is a separation between identity and clinical record, reunited only at the moment of retrieval.

How the Split Works

1. Identity is replaced with a reference. A patient becomes, say, 123. The clinical record reads "patient 123 has a headache."

2. The two halves live on separate entities. The identity map and the clinical data are stored apart — and in a hospital deployment, can be hosted locally, still in separate places.

3. A breach of one half yields very little. Break into the clinical side and you get numbers with symptoms attached to them. Break into the identity side and you get a name mapped to a number, with no medical content.

4. Retrieval rejoins them briefly, then lets go. The system pulls both together to answer the question, and they return to their separate homes immediately after.

5. Encryption sits on top of all of it. Pranav is upfront that he is simplifying: "It's much more complex than that."

The consequence that matters most: the AI learns from symptom patterns, not from identifiable people. This is a good instinct to see in an Indian health-tech company — Dozee's contactless patient monitoring ran into the same wall years earlier, that continuous clinical data is only as adoptable as the trust wrapped around it.

Equalising Medical Information

Asked about the future, Pranav goes back to the thing he says he repeats to everyone.

"AI has given us this ability to equalise information regardless of who you are or which background you come from."

His observation is that general-purpose AI has already flattened access to ordinary information — ask ChatGPT a question from a rural town or from Dubai and you get broadly the same answer. That flattening simply hasn't reached medicine, where the quality of what you get still depends heavily on where your doctor trained and what they've seen.

ZenMD's ambition is to extend it. An RMP practitioner in a small town and an MBBS doctor in a metro should both be able to ask whether a headache is just a headache, or whether there's a symptom cluster documented somewhere in Norway that changes the picture. And in reverse — a doctor in Miami should be able to consider the tapeworm, not just the lactose intolerance.

There's a second constituency they didn't fully plan for: students. The platform carries a NEET practice module that generates questions, marks the answers and points to better sources when the answer is wrong. Medical students are running mock drills on it, which puts ZenMD in the same territory as DocTutorials' medical entrance prep, but arriving from the opposite direction — a clinical tool that picked up students, rather than a student tool trying to become clinical.

The end state Pranav describes is deliberately unglamorous.

"Our goal — mine and Rithika's specifically — is to build it to a point where the doctors start using it and it doesn't feel like another login. It just stays on my second computer. I'm working on this side, I'm asking here."

Not a destination product. Ambient infrastructure. The measure of success is that nobody notices they're using it.

ZenMD by the Numbers

Launched: India, May 2026. Hyderabad- and US-based.

Adoption: more than 60,000 users and over 90,000 pageviews within eight weeks of launch, through organic adoption, per company figures.

Specialties: 14-plus covered by the validating doctor panel at the time of this conversation, targeting close to 50; company materials now cite support for more than 80 medical specialties.

What doctors actually use it for (Hyderabad usage data): diagnosis and symptom analysis 21.6%, drug information and dosing 15.6%, treatment planning and management 10.9%, lab and imaging interpretation 9.3%.

Compliance: built for HIPAA, DPDP and ABDM from day one.

Key Takeaways

Refusal is a feature, not a gap: In high-stakes domains, a system that says "I cannot verify this" is more valuable than one that always produces an answer. ZenMD's differentiator is the sentence it will say, not the ones it won't.
Agreeableness is the real hallucination risk: General assistants are tuned to confirm the user. A doctor who already has a hypothesis needs a challenger, not an echo.
You can't prompt your way out of a population mismatch: Indian drug doses, diets and prevalent conditions had to be foundational, not adapted at the edges.
Not being a domain expert is survivable, if you know which half you own: The founders own behavioural data and product; the doctors own medicine. The mistake is pretending to own both.
Ship, then find out what people actually want: The team expected documentation to lead. Drug dosing turned out to be the killer use case — and the published usage data confirmed it.
Sell time, not intelligence: In a hundred-patient day, the constraint is minutes per patient. Efficiency is a pitch a busy clinician can evaluate immediately; "smarter AI" is not.
Compliance cannot be retrofitted: Build to the strictest applicable regime from day one. You can loosen a tight architecture; you cannot tighten a leaky one after data is already out.
Separate identity from record: Storing who a patient is apart from what they have means a single breach yields either names or symptoms — and a model that learns from patterns, not people.
Track repeat usage, not signups: In a profession that runs on trust and paper, coming back is the only honest signal of adoption.

About the Guest

Pranav Reddy is the Co-Founder and Chief Marketing Officer of ZenMD. Based primarily in the US and schooled there, he has spent his career in marketing — beginning with an agency internship, rising to director, and later founding his own agency — driven throughout by an interest in human behaviour and behavioural data. He co-founded ZenMD with his sister, Rithika Reddy, the company's Founder and CEO, who brings over a decade of experience building fintech and lending platforms.

ZenMD is an India-focused clinical AI platform for doctors, clinics and hospitals, launched in India in May 2026 and operating out of Hyderabad and the United States. It offers four integrated workflows — ZenScribe for real-time voice-to-clinical documentation, ZenNotes for automated OPD notes, ZenConsult for evidence-based clinical decision support, and ZenScan for AI-powered medical imaging analysis — sitting on top of a multi-layer verification system designed to cite its sources, report its confidence level, and decline to answer when it cannot verify a claim. The platform is built for HIPAA, DPDP and ABDM compliance, with patient identity stored separately from clinical records. Company figures cite more than 60,000 users within eight weeks of launch and support for over 80 medical specialties, alongside a NEET practice module used by medical students.