We rebuilt Saath in under a day for a Google buildathon. It got us to the city round, not past it.

Carson Rodrigues

Carson Rodrigues / November 21, 2026

6 min read • ––– views

hackathonsai-agentshealth
The Saath website: 'Your health insurance, on your side', in English and Hindi

When someone you love is being admitted to hospital, the desk asks for the insurance policy number. Often nobody in the family knows where the policy is, and they end up paying out of their own pocket even though they're insured.

That is the problem behind Saath, which means "together" in Hindi.

HackathonStartup Buildathon India 2026
Organiserhack2skill, with Google DeepMind and Google Cloud
WhenIdea round sent 12 November 2026; city round 20 and 21 November
Result and rankShortlisted to the city round, not selected further. No rank or field size to quote
Built inUnder a day for the rebuild: 10:51 IST on 20 November to before 09:00 IST on 21 November
TeamTwo of us
Money spent₹614.16 billed by Google Cloud (1 to 21 November; ₹1,079.83 of usage, less ₹465.67 free tier) plus about ₹140 for the domain, about ₹754 in all
CodePrivate

The hackathon

Startup Buildathon India 2026 was run by hack2skill with Google DeepMind and Google Cloud. Round one was an idea submission, which we sent on 12 November. After we were shortlisted, the city round ran on 20 and 21 November.

We rebuilt the whole product on the organisers' Google stack in that window. We started at 10:51 IST on the 20th and submitted before the 09:00 IST deadline on the 21st. That's under a day.

The problem, in numbers

India's insurance regulator, IRDAI, publishes a yearly report. The 2024-25 report says 58 crore people are covered by health insurance. Of the claims settled, 41% went the reimbursement way: the family pays first and claims the money back later. Policyholder complaints rose 20% in a year, to 2,57,790.

We built for the people who struggle most at that desk. Older parents who can't read a 40-page policy. First-timers who have never heard the word TPA (the middleman that handles claims). The adult child who is suddenly in charge.

What Saath does

Three things matter most.

The Policy X-ray. You upload or photograph a policy once, and Saath lists the catches in plain words, like a cap on the room rent. Each one points to the page in the policy and the insurance rule behind it.

The emergency pass. Every family member gets one, with their policy number, member ID, the helpline for the claims middleman, blood group and allergies. It opens from a link or a QR code on any phone, no app needed.

Hospital Mode. At the desk, you talk to Saath in Hindi, English or other Indian languages. It already knows the family's passes and the rules, and it walks you through the desk one step at a time. The voice runs on Gemini Live, Google's live voice model. You can also photograph the admission form or a bill, and it marks what to fill and which lines are often not paid.

Flow diagram: a policy is read into a Policy X-ray using IRDAI rules; family passes feed Hospital Mode, a voice that also reads photos from the desk; other ways in are WhatsApp, email, a watch and an offline Android app

Around that we added an Android version that works with no signal, a watch app for the pass, and assistants for WhatsApp and email. Behind it sits a team of agents built with Google's agent kit, and the IRDAI rules are stored in a database that can search by meaning.

Research

Primary research was thin. I have no record of user interviews, and the policy reader was tried on sample policy documents, not on real families' papers.

Secondary research. The numbers above come from the IRDAI annual report for 2024-25: 58 crore covered, 41% of settled claims reimbursed, complaints up 20%. The same report says ₹29,933 crore of claims weren't paid out in the year. The rules Saath quotes are IRDAI's own, which we loaded and made searchable. I don't have a record of studying other insurance apps.

How I tested it

Not as hard as I'd like, and the limits matter.

  • The rules engine, which flags the catches, was checked against saved expected answers from the round-one version, so the rebuilt one had to give the same results.
  • The policy reader and the voice's decisions each had a set of golden answers, with pass marks and a live test that scores them against the real services. The live test runs only on demand. I don't have its scores in the write-up, so I won't quote one.
  • 28 automated test files on the agent side, covering the family passes, the vault, WhatsApp and email handling, the public interface and more. 139 commits and 12 merged pull requests in all.
  • Test data was sample policy documents and made-up families. Saath Insights ran on de-identified, synthetic data.

What I didn't do: no held-out set of real policies, no measured accuracy for the voice at a real desk, and no test with a user who wasn't us.

What happened

We were shortlisted for the city round. We weren't selected for the next one. I have no rank or field size to quote, only that result.

It was a one-day buildathon, but the round also included established companies showing products they had worked on for months or years. Ours was rebuilt that day, so it was never an even contest. I'm not saying that to excuse it, just to describe the room.

The judges gave honest feedback, and we wrote it down.

  • Show a business model. We had features first and a business model second.
  • WhatsApp is a strong channel. Ours only worked on Meta's test number, which you can't show to judges.
  • "Make your family hospital-ready" didn't land. It sounds like preparing for illness. We changed the line to "Your health insurance, on your side."

What we learned

Business model first. We built twelve things and then had to explain who would pay. After the feedback we wrote a short brief: free for families, with hospitals, insurers, claims middlemen and employers paying, and selling policies only later, with a licence. That should have been written before the demo.

Only show what works live. A WhatsApp assistant that needs a verified phone number isn't a channel yet. I'd rather show three things that work on stage than twelve where one can't be shown.

Lead with the moment at the desk. It's what people remember. The features came after, and I think they buried it.

Links

Once the round ended we shut everything down. The cloud services and all family data were deleted on 21 November. The code is in a private repository, so there's nothing to link there. A static page is still up.

Available · London / UK · senior AI, AI FDE, contract

Building something with AI?

Voice agents, MCP servers, LLM pipelines, agentic workflows — pick a slot, drop a message, or send your email and I'll reply within a day.

or leave your email

Replies within ~24 hours · London, UK · on-site, hybrid or remote