India Case Study.

Large Context Imports & Model Optimization

Spent today importing massive patient histories into Vivral. The context sizes we were dealing with were huge. To make it work smoothly, we actually had to go in and rewrite the core Patient Passport backend so it could handle and search through that much data without choking.

We also realized that when people ask simple, casual questions (like "how do I bathe a patient who can't move?"), the AI was giving pretty weak answers compared to when we asked long, formal medical questions. So, we've started some Reinforcement Learning (RL) to train the model to be better at handling just regular, everyday questions.

Wearable Integration & Information Refinement

The main goal today was testing Apple Health imports for user-authorized longitudinal data, such as overnight vitals and sleep summaries. Longer-term context can make educational summaries more useful, but it does not establish diagnostic accuracy and must be checked against the original record and professional care.

We also pushed some patches to fix latency when the chat gets really long. While testing, we caught the AI hallucinating some weird stuff about PSA counts. Looked into it and found out it was just a typo error when it was reading the data. We patched it and we're hitting it with some RLHF to make sure it doesn't happen again.

User Interaction & Model Preference

We're noticing the patient and their caretaker are using Vivral a lot—like 10 to 20 questions a day, asking everything from basic care stuff to deep medical questions. Vivral Swift is definitely the favorite model right now just because it's so fast and handles almost everything they throw at it.

Honestly, we haven't seen a real need to switch them to the Surreal model yet. Right now, we're shifting our focus to polish up the rehab features in the app to help out more with physical recovery.

Hospital Logistics & Telemedicine Validation

Going to the hospital today really showed us the logistical nightmare these patients face. The place was insanely crowded, and we couldn't even find a wheelchair, which made getting around super tough and physically draining for the patient.

After a long wait, a clinician evaluated the patient and reached a diagnosis. An earlier Vivral response had mentioned a similar possibility, but it could not confirm the condition and was not a substitute for the examination. The experience reinforced the need for better care access—not the idea that an AI response can replace a visit.

Patient Passport & Automated Care Management

Today we really pushed the Patient Passport feature. We used Vivral's OCR to scan and upload our patient's entire medical history. It ate up all those complex files without a problem. Almost right away, Vivral started using that context to send automated, proactive notifications, like reminding him about upcoming follow-up visits.

It was amazing to see. We realized that translating confusing doctor's notes and discharge papers into plain English is huge—especially for older patients who might be forgetful after a big hospital stay. When they go home, they usually don't have the energy to manage their own care, so having an app automatically handle the schedule means important steps aren't forgotten.

Data Tracking & Supervised Care Adjustments

A high blood-glucose reading prompted the care team to contact the patient's licensed clinician. Vivral was used only to organize readings; no insulin or other treatment change was made from an App output.

It's really important to note that this whole process was strictly supervised and approved by a licensed doctor. Vivral was just the tool keeping everything organized and tracking the data so the doctor could make the right call. The patient's blood sugar is stable now, and Vivral is just running in the background, keeping a clean log of his vitals.

Systemic Burden & The Need for Triage

I was talking with a patient's sibling today, and she told me a really heartbreaking story about losing her husband. He had a poorly done surgery, was in immense pain, and was rushed to the hospital showing stroke symptoms. But the hospital was so overcrowded that no doctor was available to see him. They waited for hours and then tried to transfer him somewhere else, but he passed away before anyone could help.

This story just proves why Vivral is so necessary. When we visit these hospitals, they are packed to the brim, a lot of times with people who don't even need acute emergency care. If we can use Vivral to handle the routine cases and do remote triage, we can clear out the waiting rooms so real doctors actually have time for the life-threatening emergencies.

Rural Connectivity and the Triage Gap

Today we took a trip to a more northern region in Kerala. The internet out there was so terrible in so many areas that even Google Maps didn't work. We passed by thousands of villages, and the nearest functional hospital was over 100km away.

For these thousands of people, they all have cell phones, but they just don't have access to credible medical care on a daily basis. Most of these people don't even have usable cars. It's either they suffer and let NGOs and volunteer physicians occasionally take care of them, or they attempt a massive journey without knowing if they actually need to.

With an AI triage layer like Vivral, they can get a quick look at their symptoms right on their phone to see the severity of their case and figure out whether that 100km hospital trip is actually worth it. We clearly found a massive place for Vivral to be used out here.

Overcoming Literacy and Language Barriers

As we're rolling this out to more people, it's pretty clear that a lot of patients couldn't speak English well, and many couldn't write entirely. An AI assistant is basically useless if the patient can't even talk to it.

To fix this fast, we added a lightweight translation and transcription layer right into the app. Now, patients can just speak to Vivral in their local language without needing to type anything. Vivral takes the audio, translates it, figures out what they need, and speaks the clinical response back to them out loud. By making it voice-first, we completely bypassed the reading and writing barrier. It's a huge step in making sure the people who need it most can actually use it.

Experimental Camera Pulse Observation

During historical supervised research, the team compared an experimental camera pulse estimate with a manual pulse count. The participant had an implanted pacemaker, which made it especially important not to treat a phone result as evidence about the device or its battery.

The experimental outputs were observational only and were checked by a nurse. They were not used to monitor the pacemaker, assess danger, decide whether to exercise or travel, or change the replacement plan. Only the treating cardiac team and validated equipment could assess those questions.

The current feature is called Pulse Wellness Check. It is a general-wellness camera estimate—not a heart-rate monitor, medical device, continuous monitor, or tool for evaluating a pacemaker or other implanted device. Users should verify important readings with an appropriate validated device or clinician.

Visual Information Limits & Severe Hospital Overcrowding

Before a checkup, the team noticed dark spots on the patient's arm. An experimental image response surfaced several possibilities and substantial uncertainty. It could not identify a clot or determine urgency; later clinical evaluation supplied the medically meaningful assessment.

When we went to the hospital, where his doctor is, the doctor didn't show up for hours, and there were 32 patients in front of him. Out of frustration and out of tiredness, our patient chose to go home and rest. He rescheduled for a day where the queue was certain to be smaller. This just highlights how poor the healthcare system is over here. This is the best hospital within a reasonable radius and they are extremely overworked.

This episode exposed an important model limitation: image outputs could be alarming, over-broad, or wrong. It became a safety and evaluation case for later model work, not evidence that Vivral can perform visual diagnosis.