An engineer who has built, shipped and sold AI products before.

Aryan Kumar · IIT Madras · Lead AI Engineer in Japan

Role: Lead AI Engineer in Japan, shipping production AI systems that regulated businesses depend on. Previous exit: built Donna, a personal AI assistant, from scratch and sold it to SpringCube.

Research: early detection of Parkinson's and Alzheimer's, which became NeuroConnect, a community diagnostic app presented at INSEAD Paris. That research is the thesis: in aging, the signal is behavioural and early, long before it is clinical.

CareOS, built from scratch, entirely solo: the 17-language voice agent, streaming telephony, mid-call emergency detection, backend, security, two web apps, a mobile app, and live billing. Sub-second voice, encrypted health data, 484 automated tests, live in production with paying customers.

He is also the customer: an immigrant in Osaka with parents in another country. Backed by the startup programs of Anthropic, OpenAI, AWS and Deepgram.

In aging, the signal is behavioural, gradual, and daily.

Both are asking the same question every day: is my person okay today?

Voice AI crossed the threshold where a phone call stops feeling like a machine.

Under one second it reads as conversation; at 25 seconds an elderly person assumes the line is dead and hangs up. And this now holds in Hindi, Tamil and Bengali, not just English. That is the unlock for every market outside the West, and it is very recent.

The capital already moved. It went to operations, not companionship.

Every one of the eight largest rounds sells to agencies, facilities or payers. The only pure consumer-companion rounds were an order of magnitude smaller. That is exactly why CareOS sells B2B first and families second.

The care manager opens one screen and knows who to worry about.

A risk queue in the order a care manager actually works. Every call transcribed, summarised, scored, and drafted into a record awaiting approval.

Every patient, every call, every medication. One searchable place.

The paper file, the phone log and the care manager's memory, replaced by one screen that never forgets. This is the memory the record is drafted from.

A conversation becomes a checklist nobody has to remember.

The parent's consent, quoted in her own words, waiting on the child's one tap. The app never silently forgets something a parent asked for.

It doesn't just listen. With consent, it acts, during the call.

The parent agrees on the phone → the agent files it and finds options while they wait → the family gets the findings and one tap releases it. They can chat with CareOS about any task. Money never moves: payment-shaped requests are refused in code, not in a prompt.

For the child who left, and still lies awake about it.

Read what your mother actually said, not a summary of it. Approve what the agent proposes. Be told the moment something is wrong.

Three care organisations. Twelve paying customers. Zero spend.

I started building for Japan. None of my first partners is in Japan. That is what made me stop building a Japanese product.

$5 for the first 50 families. $15 for the first 10 agencies.

Pay with your email and that email becomes your login. Checkout, entitlement, and app access are one wired loop, with the pilot caps enforced in the database, not on the landing page.

Five signals, none of them bought.

304 million people live in a different country from their parents.

Migration is 3.7% of humanity and rising. Every one of those people has parents somewhere else, and no way to know how today went.

We are not sizing a software category. We are sizing the work being done unpaid.

The $1.01 trillion is the ceiling: work families already do for free, which exceeds all US Medicaid spending combined. The paid lines beneath it are what is already transacting. The smallest of them is where CareOS bills today.

From three organisations to the layer both sides depend on.

Bottom-up from live pricing, not a share-of-market estimate. The seven research firms sizing 'home health software' disagree by 4.9×, so we do not use their numbers.

Three revenue lines, and one patient can pay on all three.

A family on $79 whose mother is served by a partnered agency generates ~$110/month from two payers, neither paying full freight.

Where this reaches $100M, and why it does not stop there.

Scale needs 70,000 paying families worldwide: one tenth of one percent of the 63 million unpaid family caregivers in the US alone, before counting India, Japan or the Gulf.

Where reimbursement is capped, software is revenue.

QuroCare, GeroCare and Home Healthcare agreed to $39 because the maths worked, not because the demo was charming.

Everyone stops at the summary. The record is where the money is.

The first steps are a commodity: four companies shipped them this year. The last two are national regulation, and they are why an agency can adopt us at all and why a family gets something they can act on.

Everyone sells the conversation. We sell the record.

The field is real and funded: Hippocratic AI ($3.5B), Kaigo Health (YC F25), inTouch, and in Japan Welmo (¥5.6B raised, already shipping care-plan AI to our buyer). And 'no app for the elder' is table stakes, not a moat: at least eight companies say it.

Live in production today, not a prototype.

Reviewers can log into both products today and place a real call.

Built like the health product it is, not the demo it started as.

484 automated tests run on every change. The same discipline the founder ships professionally, applied to his own company.

The map is drawn. The constraint is nights and weekends.

Everything on this list extends what is already running in production. None of it is greenfield.