Taking agentic AI healthcare from 0 to 1, for doctors and for patients.

The bet

Imagine every person, wherever they live and whatever language they speak, having a personal doctor available 24x7. One who knows their history, keeps track of their health between visits, and connects them to the right care when they need it.

That is the bet behind Pranik.ai, an AI health-tech venture started under the Tabhi group alongside Mondee. Care in India today is mostly episodic. You fall sick, you queue at an OPD, a doctor who has a few minutes for you writes a prescription, and the system forgets you until next time. The doctor, meanwhile, spends a large part of each consult typing.

The bet

From episodes to a loop

Today · episodic

1. 01Fall sick
2. 02Queue at the OPD
3. 03A few minutes with a doctor
4. 04Paper prescription
5. 05Forgotten until next time

With Pranik · continuous

1. 01 · PeopleAsk anytime, in your language
2. 02 · PeopleMatched to the right doctor
3. 03 · DoctorsConsult while the AI listens and drafts
4. 04 · PeopleFollow-ups, reminders, questions answered
5. 05 · BothHistory carried into the next visit

care continues between visits

Agentic AI that is voice-first and multilingual could change both sides of that picture. That is what I joined to build.

Where I came in

I joined in September 2025 and started in the founder’s office on strategy, then moved into product ownership across both apps, as the team’s first product person.

In a startup that means wearing whichever hat the product needs that week. I have been the product manager, the UX researcher, the product owner, sometimes the designer, and sometimes the person pitching to a new hospital.

Two products, one marketplace

Pranik for Doctors

• AI clinical assistantListens to the consultation, drafts the prescription and case sheet, reads lab reports. Stored to EMR/EHR standards.
• Virtual clinicQueue, appointments, billing, pharmacy and online consults for doctors running their own practice.
• Chief-of-staff agentLeads a set of AI workforce agents: finance, medical research, patient follow-up and telephony, some still in testing.
• Digital twinThe doctor’s avatar and voice clone, trained on how that doctor practices.

Online consultation

two-sided marketplace

Pranik for People

• 24/7 health assistantAvatar-based and voice-first: symptom intake, appointment booking, post-visit care.
• Care continuityAfter a visit, patients can video-consult or ask the doctor’s AI assistant when the doctor has no time.
• Agentic follow-upsReminders, follow-ups and tool calls that act on the patient’s behalf.

Pranik for Doctors started as an AI clinical assistant for hospitals. It records the whole doctor-patient conversation, drafts the prescription and case sheet, offers suggestions and analyzes uploaded lab reports. Everything is stored to EMR and EHR standards, including into a hospital’s existing system. It then grew into a complete AI-powered virtual clinic for doctors who run their own practice: queues, appointments, billing, pharmacy integrations and online consultations, with AI workforce agents led by a chief-of-staff agent, some of them still in testing.

Pranik for People is a 24/7 agentic AI health assistant for sick care and post-consultation care. It is avatar-based and voice-first. It handles symptom intake, appointment booking and post-visit care, and it closes the continuity-of-care loop: after an in-person visit, a patient can video-consult or ask their doctor’s AI assistant the questions they forgot to ask.

Setting up the product function

There was no product function to inherit, so I built one:

• Wrote the product team’s SOPs, including how we write flowcharts and PRDs.
• Hired the team’s designer and mentored her, and mentored the incoming business analysts.
• Led two cross-functional squads, each with one or two mobile developers, an AI engineer and one or two full-stack or backend engineers, mentored by our senior leads. A two-person design team and a market researcher work with me.
• Ran daily standups and prioritization across three competing inputs: new features, leadership requests and what user testing told us.

Product lanes were defined with the founders, the head of engineering and the head of AI in one room. Early on I did not have the full say. Now I have more of it.

Defining what “working” means

An AI that drafts a prescription is only useful if a busy doctor keeps using it next week. So the question I kept coming back to was not “is the model accurate?” but “will this doctor still be using it a month from now, and why?”

What “working” meant for the doctor product

The outcome

Doctors keep using it in their own OPD, and bring colleagues

• ~90%of active doctors stayed for phase two
• 2xregistrations, from their referrals

Driver 01

It gives them time back

• 10+ → under 2 mindocumentation per consult at Bhaktivedanta

Driver 02

They can trust the draft

• ~70%fewer drug-name errors, from four fixes shipped together
• Blindmonthly doctor checks on what changed

Driver 03

It keeps up with the room

• ~35 saverage prescription latency, down from 2+ min

From pilots to phase two

The doctor product went into real pilots at hospitals, clinics, government health drives and medical camps. That is where the 20,000+ consultations come from.

Phase one taught us that adoption breaks on small, clinic-specific things. Doctors dropped off one by one on blockers that only existed in their setting, and waited for us to fix them.

Phase one to phase two

1. 01 · Pilot Real OPDs, camps and drives 20,000+ consultations with 100+ doctors.
2. 02 · Listen Clinic-specific blockers Feature requests, bug reports and what I saw on site.
3. 03 · Rebuild A new version from that list UI and UX, AI quality, data pipelines, RAG for medicine recommendations, the mobile app.
4. 04 · Retain ~90% of active doctors continued
5. 05 · Spread Referrals doubled registrations Inside their hospitals and outside.

At Bhaktivedanta Hospital in Mumbai, doctors used to type the full record after each consultation, which took 10+ minutes. Now they review and correct the AI’s case sheet in under 2 minutes. Bhaktivedanta went from one doctor to an entire department.

Taking AI medical software through the regulator

An AI that drafts prescriptions is medical device software, so it needs a license. I oversaw the audit, clinical validation and the ~180-page dossier behind Pranik’s CDSCO Class A test license, guided by a consultant agency and with another analyst coordinating day to day.

The evidence trail behind the license

1. 01 · Audit Organization-wide audit
2. 02 · Document Every process and clinical validation, written down
3. 03 · Trial 2,500-consultation clinical trial
4. 04 · Sign-off 25 specialists signed letters On workflow fit, error rates and time saved.
5. 05 · Dossier ~180 pages
6. 06 · License CDSCO Class A test license

I also worked on ABDM Milestone 1, the national digital health integration. My part in the license was making sure the evidence was there.

What I’m taking forward

• Adoption is local. The product that works in one OPD stalls in the next for reasons no roadmap predicts. Being on site is not optional.
• Agents need their own spec format. Traditional PRDs could not keep up, so I built one. More in the Agent Flow Blueprint case study.
• Safety beats volume. The evals loop ranks errors by clinical risk, not by how often they show up.

Next on the roadmap, the marketplace will expand to pharmacies, diagnostic labs and other services through a franchise model. That part is still ahead of us.